restructure project

2026-01-27 23:58:17 +01:00
parent 58bf75db68
commit d6550375b0
230 changed files with 5513 additions and 1756 deletions
--- a/.claude/settings.local.json
+++ b/.claude/settings.local.json
@@ -87,7 +87,10 @@
      "Bash(wsl bash -c \"cd /mnt/c/Users/yaoji/git/ColaCoder/invoice-master-poc-v2 && source ~/miniconda3/etc/profile.d/conda.sh && conda activate invoice-py311 && python -m pytest tests/ -v --tb=short 2>&1 | tail -60\")",
      "Bash(wsl bash -c \"source ~/miniconda3/etc/profile.d/conda.sh && conda activate invoice-py311 && cd /mnt/c/Users/yaoji/git/ColaCoder/invoice-master-poc-v2 && python -m pytest tests/data/test_admin_models_v2.py -v 2>&1 | head -100\")",
      "Bash(dir src\\\\web\\\\*admin* src\\\\web\\\\*batch*)",
-      "Bash(wsl bash -c \"source ~/miniconda3/etc/profile.d/conda.sh && conda activate invoice-py311 && cd /mnt/c/Users/yaoji/git/ColaCoder/invoice-master-poc-v2 && python3 -c \"\"\n# Test FastAPI Form parsing behavior\nfrom fastapi import Form\nfrom typing import Annotated\n\n# Simulate what happens when data={''upload_source'': ''ui''} is sent\n# and async_mode is not in the data\nprint\\(''Test 1: async_mode not provided, default should be True''\\)\nprint\\(''Expected: True''\\)\n\n# In FastAPI, when Form has a default, it will use that default if not provided\n# But we need to verify this is actually happening\n\"\"\")"
+      "Bash(wsl bash -c \"source ~/miniconda3/etc/profile.d/conda.sh && conda activate invoice-py311 && cd /mnt/c/Users/yaoji/git/ColaCoder/invoice-master-poc-v2 && python3 -c \"\"\n# Test FastAPI Form parsing behavior\nfrom fastapi import Form\nfrom typing import Annotated\n\n# Simulate what happens when data={''upload_source'': ''ui''} is sent\n# and async_mode is not in the data\nprint\\(''Test 1: async_mode not provided, default should be True''\\)\nprint\\(''Expected: True''\\)\n\n# In FastAPI, when Form has a default, it will use that default if not provided\n# But we need to verify this is actually happening\n\"\"\")",
+      "Bash(wsl bash -c \"cd /mnt/c/Users/yaoji/git/ColaCoder/invoice-master-poc-v2 && sed -i ''s/from src\\\\.data import AutoLabelReport/from training.data.autolabel_report import AutoLabelReport/g'' packages/training/training/processing/autolabel_tasks.py && sed -i ''s/from src\\\\.processing\\\\.autolabel_tasks/from training.processing.autolabel_tasks/g'' packages/inference/inference/web/services/db_autolabel.py\")",
+      "Bash(wsl bash -c \"source ~/miniconda3/etc/profile.d/conda.sh && conda activate invoice-py311 && cd /mnt/c/Users/yaoji/git/ColaCoder/invoice-master-poc-v2 && pytest tests/web/test_dataset_routes.py -v --tb=short 2>&1 | tail -20\")",
+      "Bash(wsl bash -c \"source ~/miniconda3/etc/profile.d/conda.sh && conda activate invoice-py311 && cd /mnt/c/Users/yaoji/git/ColaCoder/invoice-master-poc-v2 && pytest --tb=short -q 2>&1 | tail -5\")"
    ],
    "deny": [],
    "ask": [],
--- a/.coverage
+++ b/.coverage
--- a/.gitignore
+++ b/.gitignore
@@ -52,6 +52,10 @@ reports/*.jsonl
 logs/
 *.log

+# Coverage
+htmlcov/
+.coverage
+
 # Jupyter
 .ipynb_checkpoints/

--- a/README.md
+++ b/README.md
@@ -8,7 +8,25 @@

 1. **自动标注**: 利用已有 CSV 结构化数据 + OCR 自动生成 YOLO 训练标注
 2. **模型训练**: 使用 YOLOv11 训练字段检测模型
-3. **推理提取**: 检测字段区域 → OCR 提取文本 → 字段规范化
+3. **推理提取**: 检测字段区域 -> OCR 提取文本 -> 字段规范化
+
+### 架构
+
+项目采用 **monorepo + 三包分离** 架构，训练和推理可独立部署：
+
+```
+packages/
+├── shared/      # 共享库 (PDF, OCR, 规范化, 匹配, 工具)
+├── training/    # 训练服务 (GPU, 按需启动)
+└── inference/   # 推理服务 (常驻运行)
+```
+
+| 服务 | 部署目标 | GPU | 生命周期 |
+|------|---------|-----|---------|
+| **Inference** | Azure App Service | 可选 | 常驻 7x24 |
+| **Training** | Azure ACI | 必需 | 按需启动/销毁 |
+
+两个服务通过共享 PostgreSQL 数据库通信。推理服务通过 API 触发训练任务，训练服务从数据库拾取任务执行。

 ### 当前进度

@@ -16,6 +34,8 @@
 |------|------|
 | **已标注文档** | 9,738 (9,709 成功) |
 | **总体字段匹配率** | 94.8% (82,604/87,121) |
+| **测试** | 922 passed |
+| **模型 mAP@0.5** | 93.5% |

 **各字段匹配率:**

@@ -42,24 +62,83 @@
 |------|------|
 | **WSL** | WSL 2 + Ubuntu 22.04 |
 | **Conda** | Miniconda 或 Anaconda |
-| **Python** | 3.10+ (通过 Conda 管理) |
+| **Python** | 3.11+ (通过 Conda 管理) |
 | **GPU** | NVIDIA GPU + CUDA 12.x (强烈推荐) |
 | **数据库** | PostgreSQL (存储标注结果) |

-## 功能特点
+## 安装

- **双模式 PDF 处理**: 支持文本层 PDF 和扫描图 PDF
- **自动标注**: 利用已有 CSV 结构化数据自动生成 YOLO 训练数据
- **多策略字段匹配**: 精确匹配、子串匹配、规范化匹配
- **数据库存储**: 标注结果存储在 PostgreSQL，支持增量处理和断点续传
- **YOLO 检测**: 使用 YOLOv11 检测发票字段区域
- **OCR 识别**: 使用 PaddleOCR v5 提取检测区域的文本
- **统一解析器**: payment_line 和 customer_number 采用独立解析器模块
- **交叉验证**: payment_line 数据与单独检测字段交叉验证，优先采用 payment_line 值
- **文档类型识别**: 自动区分 invoice (有 payment_line) 和 letter (无 payment_line)
- **Web 应用**: 提供 REST API 和可视化界面
- **增量训练**: 支持在已训练模型基础上继续训练
- **内存优化**: 支持低内存模式训练 (--low-memory)
+```bash
+# 1. 进入 WSL
+wsl -d Ubuntu-22.04
+
+# 2. 创建 Conda 环境
+conda create -n invoice-py311 python=3.11 -y
+conda activate invoice-py311
+
+# 3. 进入项目目录
+cd /mnt/c/Users/yaoji/git/ColaCoder/invoice-master-poc-v2
+
+# 4. 安装三个包 (editable mode)
+pip install -e packages/shared
+pip install -e packages/training
+pip install -e packages/inference
+```
+
+## 项目结构
+
+```
+invoice-master-poc-v2/
+├── packages/
+│   ├── shared/                     # 共享库
+│   │   ├── setup.py
+│   │   └── shared/
+│   │       ├── pdf/                # PDF 处理 (提取, 渲染, 检测)
+│   │       ├── ocr/                # PaddleOCR 封装 + 机器码解析
+│   │       ├── normalize/          # 字段规范化 (10 种 normalizer)
+│   │       ├── matcher/            # 字段匹配 (精确/子串/模糊)
+│   │       ├── utils/              # 工具 (验证, 清理, 模糊匹配)
+│   │       ├── data/               # DocumentDB, CSVLoader
+│   │       ├── config.py           # 全局配置 (数据库, 路径, DPI)
+│   │       └── exceptions.py       # 异常定义
+│   │
+│   ├── training/                   # 训练服务 (GPU, 按需)
+│   │   ├── setup.py
+│   │   ├── Dockerfile
+│   │   ├── run_training.py         # 入口 (--task-id 或 --poll)
+│   │   └── training/
+│   │       ├── cli/                # train, autolabel, analyze_*, validate
+│   │       ├── yolo/               # db_dataset, annotation_generator
+│   │       ├── processing/         # CPU/GPU worker pool, task dispatcher
+│   │       └── data/               # training_db, autolabel_report
+│   │
+│   └── inference/                  # 推理服务 (常驻)
+│       ├── setup.py
+│       ├── Dockerfile
+│       ├── run_server.py           # Web 服务器入口
+│       └── inference/
+│           ├── cli/                # infer, serve
+│           ├── pipeline/           # YOLO 检测, 字段提取, 解析器
+│           ├── web/                # FastAPI 应用
+│           │   ├── api/v1/         # REST API (admin, public, batch)
+│           │   ├── schemas/        # Pydantic 数据模型
+│           │   ├── services/       # 业务逻辑
+│           │   ├── core/           # 认证, 调度器, 限流
+│           │   └── workers/        # 后台任务队列
+│           ├── validation/         # LLM 验证器
+│           ├── data/               # AdminDB, AsyncRequestDB, Models
+│           └── azure/              # ACI 训练触发器
+│
+├── migrations/                     # 数据库迁移
+│   ├── 001_async_tables.sql
+│   ├── 002_nullable_admin_token.sql
+│   └── 003_training_tasks.sql
+├── frontend/                       # React 前端 (Vite + TypeScript)
+├── tests/                          # 测试 (922 tests)
+├── docker-compose.yml              # 本地开发 (postgres + inference + training)
+├── run_server.py                   # 快捷启动脚本
+└── runs/train/                     # 训练输出 (weights, curves)
+```

 ## 支持的字段

@@ -76,476 +155,129 @@
 | 8 | payment_line | 支付行 (机器可读格式) |
 | 9 | customer_number | 客户编号 |

-## DPI 配置
-
-**重要**: 系统所有组件统一使用 **150 DPI**，确保训练和推理的一致性。
-
-DPI（每英寸点数）设置必须在训练和推理时保持一致，否则会导致：
- 检测框尺寸失配
- mAP显著下降（可能从93.5%降到60-70%）
- 字段漏检或误检
-
-### 配置位置
-
-| 组件 | 配置文件 | 配置项 |
-|------|---------|--------|
-| **全局常量** | `src/config.py` | `DEFAULT_DPI = 150` |
-| **Web推理** | `src/web/config.py` | `ModelConfig.dpi` (导入自 `src.config`) |
-| **CLI推理** | `src/cli/infer.py` | `--dpi` 默认值 = `DEFAULT_DPI` |
-| **自动标注** | `src/config.py` | `AUTOLABEL['dpi'] = DEFAULT_DPI` |
-| **PDF转图** | `src/web/api/v1/admin/documents.py` | 使用 `DEFAULT_DPI` |
-
-### 使用示例
-
-```bash
-# 训练（使用默认150 DPI）
-python -m src.cli.autolabel --dual-pool --cpu-workers 3 --gpu-workers 1
-
-# 推理（默认150 DPI，与训练一致）
-python -m src.cli.infer -m runs/train/invoice_fields/weights/best.pt -i invoice.pdf
-
-# 手动指定DPI（仅当需要与非默认训练DPI的模型配合时）
-python -m src.cli.infer -m custom_model.pt -i invoice.pdf --dpi 150
-```
-
-## 安装
-
-```bash
-# 1. 进入 WSL
-wsl -d Ubuntu-22.04
-
-# 2. 创建 Conda 环境
-conda create -n invoice-py311 python=3.11 -y
-conda activate invoice-py311
-
-# 3. 进入项目目录
-cd /mnt/c/Users/yaoji/git/ColaCoder/invoice-master-poc-v2
-
-# 4. 安装依赖
-pip install -r requirements.txt
-
-# 5. 安装 Web 依赖
-pip install uvicorn fastapi python-multipart pydantic
-```
-
 ## 快速开始

-### 1. 准备数据
-
-```
-~/invoice-data/
-├── raw_pdfs/
-│   ├── {DocumentId}.pdf
-│   └── ...
-├── structured_data/
-│   └── document_export_YYYYMMDD.csv
-└── dataset/
-    └── temp/  (渲染的图片)
-```
-
-CSV 格式:
-```csv
-DocumentId,InvoiceDate,InvoiceNumber,InvoiceDueDate,OCR,Bankgiro,Plusgiro,Amount
-3be53fd7-...,2025-12-13,100017500321,2026-01-03,100017500321,53939484,,114
-```
-
-### 2. 自动标注
+### 1. 自动标注

 ```bash
 # 使用双池模式 (CPU + GPU)
-python -m src.cli.autolabel \
+python -m training.cli.autolabel \
    --dual-pool \
    --cpu-workers 3 \
    --gpu-workers 1

 # 单线程模式
-python -m src.cli.autolabel --workers 4
+python -m training.cli.autolabel --workers 4
 ```

-### 3. 训练模型
+### 2. 训练模型

 ```bash
 # 从预训练模型开始训练
-python -m src.cli.train \
+python -m training.cli.train \
    --model yolo11n.pt \
    --epochs 100 \
    --batch 16 \
    --name invoice_fields \
    --dpi 150

-# 低内存模式 (适用于内存不足场景)
-python -m src.cli.train \
+# 低内存模式
+python -m training.cli.train \
    --model yolo11n.pt \
    --epochs 100 \
    --name invoice_fields \
-    --low-memory \
-    --workers 4 \
-    --no-cache
+    --low-memory

-# 从检查点恢复训练 (训练中断后)
-python -m src.cli.train \
+# 从检查点恢复训练
+python -m training.cli.train \
    --model runs/train/invoice_fields/weights/last.pt \
    --epochs 100 \
    --name invoice_fields \
    --resume
 ```

-### 4. 增量训练
-
-当添加新数据后，可以在已训练模型基础上继续训练：
-
-```bash
-# 从已训练的 best.pt 继续训练
-python -m src.cli.train \
-    --model runs/train/invoice_yolo11n_full/weights/best.pt \
-    --epochs 30 \
-    --batch 16 \
-    --name invoice_yolo11n_v2 \
-    --dpi 150
-```
-
-**增量训练建议**:
-
-| 场景 | 建议 |
-|------|------|
-| 添加少量新数据 (<20%) | 继续训练 10-30 epochs |
-| 添加大量新数据 (>50%) | 继续训练 50-100 epochs |
-| 修正大量标注错误 | 从头训练 |
-| 添加新的字段类型 | 从头训练 |
-
-### 5. 推理
+### 3. 推理

 ```bash
 # 命令行推理
-python -m src.cli.infer \
+python -m inference.cli.infer \
    --model runs/train/invoice_fields/weights/best.pt \
    --input path/to/invoice.pdf \
    --output result.json \
    --gpu
-
-# 批量推理
-python -m src.cli.infer \
-    --model runs/train/invoice_fields/weights/best.pt \
-    --input invoices/*.pdf \
-    --output results/ \
-    --gpu
 ```

-**推理结果包含**:
- `fields`: 提取的字段值 (InvoiceNumber, Amount, payment_line, customer_number 等)
- `confidence`: 各字段的置信度
- `document_type`: 文档类型 ("invoice" 或 "letter")
- `cross_validation`: payment_line 交叉验证结果 (如果有)
-
-### 6. Web 应用
-
-**在 WSL 环境中启动**:
+### 4. Web 应用

 ```bash
-# 方法 1: 从 Windows PowerShell 启动 (推荐)
+# 从 Windows PowerShell 启动
 wsl bash -c "source ~/miniconda3/etc/profile.d/conda.sh && conda activate invoice-py311 && cd /mnt/c/Users/yaoji/git/ColaCoder/invoice-master-poc-v2 && python run_server.py --port 8000"

-# 方法 2: 在 WSL 内启动
-conda activate invoice-py311
-cd /mnt/c/Users/yaoji/git/ColaCoder/invoice-master-poc-v2
-python run_server.py --port 8000
-
-# 方法 3: 使用启动脚本
-./start_web.sh
+# 启动前端
+cd frontend && npm install && npm run dev
+# 访问 http://localhost:5173
 ```

-**服务启动后**:
- 访问 **http://localhost:8000** 使用 Web 界面
- 服务会自动加载模型 `runs/train/invoice_fields/weights/best.pt`
- GPU 默认启用，置信度阈值 0.5
+### 5. Docker 本地开发

-#### Web API 端点
+```bash
+docker-compose up
+# inference: http://localhost:8000
+# training: 轮询模式自动拾取任务
+```
+
+## 训练触发流程
+
+推理服务通过 API 触发训练，训练在独立的 GPU 实例上执行：
+
+```
+Inference API                    PostgreSQL              Training (ACI)
+    |                                |                        |
+    POST /admin/training/trigger     |                        |
+    |-> INSERT training_tasks ------>| status=pending          |
+    |-> Azure SDK: create ACI --------------------------------> 启动
+    |                                |                        |
+    |                                |<-- SELECT pending -----+
+    |                                |--- UPDATE running -----+
+    |                                |                   执行训练...
+    |                                |<-- UPDATE completed ---+
+    |                                |    + model_path        |
+    |                                |    + metrics      自动关机
+    |                                |                        |
+    GET /admin/training/{id}         |                        |
+    |-> SELECT training_tasks ------>|                        |
+    +-- return status + metrics      |                        |
+```
+
+## Web API 端点
+
+**Public API:**

 | 方法 | 端点 | 描述 |
 |------|------|------|
-| GET | `/` | Web UI 界面 |
 | GET | `/api/v1/health` | 健康检查 |
 | POST | `/api/v1/infer` | 上传文件并推理 |
 | GET | `/api/v1/results/{filename}` | 获取可视化图片 |
+| POST | `/api/v1/async/infer` | 异步推理 |
+| GET | `/api/v1/async/status/{task_id}` | 查询异步任务状态 |

-#### API 响应格式
+**Admin API** (需要 `X-Admin-Token` header):

-```json
-{
-  "status": "success",
-  "result": {
-    "document_id": "abc123",
-    "document_type": "invoice",
-    "fields": {
-      "InvoiceNumber": "12345",
-      "Amount": "1234.56",
-      "payment_line": "# 94228110015950070 # > 48666036#14#",
-      "customer_number": "UMJ 436-R"
-    },
-    "confidence": {
-      "InvoiceNumber": 0.95,
-      "Amount": 0.92
-    },
-    "cross_validation": {
-      "is_valid": true,
-      "ocr_match": true,
-      "amount_match": true
-    }
-  }
-}
-```
-
-## 训练配置
-
-### YOLO 训练参数
-
-```bash
-python -m src.cli.train [OPTIONS]
-
-Options:
-  --model, -m        基础模型 (默认: yolo11n.pt)
-  --epochs, -e       训练轮数 (默认: 100)
-  --batch, -b        批大小 (默认: 16)
-  --imgsz            图像尺寸 (默认: 1280)
-  --dpi              PDF 渲染 DPI (默认: 150)
-  --name             训练名称
-  --limit            限制文档数 (用于测试)
-  --device           设备 (0=GPU, cpu)
-  --resume           从检查点恢复训练
-  --low-memory       启用低内存模式 (batch=8, workers=4, no-cache)
-  --workers          数据加载 worker 数 (默认: 8)
-  --cache            缓存图像到内存
-```
-
-### 训练最佳实践
-
-1. **禁用翻转增强** (文本检测):
-   ```python
-   fliplr=0.0, flipud=0.0
-   ```
-
-2. **使用 Early Stopping**:
-   ```python
-   patience=20
-   ```
-
-3. **启用 AMP** (混合精度训练):
-   ```python
-   amp=True
-   ```
-
-4. **保存检查点**:
-   ```python
-   save_period=10
-   ```
-
-### 训练结果示例
-
-**最新训练结果** (100 epochs, 2026-01-22):
-
-| 指标 | 值 |
-|------|-----|
-| **mAP@0.5** | 93.5% |
-| **mAP@0.5-0.95** | 83.0% |
-| **训练集** | ~10,000 张标注图片 |
-| **字段类型** | 10 个字段 (新增 payment_line, customer_number) |
-| **模型位置** | `runs/train/invoice_fields/weights/best.pt` |
-
-**各字段检测性能**:
- 发票基础信息 (InvoiceNumber, InvoiceDate, InvoiceDueDate): >95% mAP
- 支付信息 (OCR, Bankgiro, Plusgiro, Amount): >90% mAP
- 组织信息 (supplier_org_number, customer_number): >85% mAP
- 支付行 (payment_line): >80% mAP
-
-**模型文件**:
-```
-runs/train/invoice_fields/weights/
-├── best.pt          # 最佳模型 (mAP@0.5 最高) ⭐ 推荐用于生产
-└── last.pt          # 最后检查点 (用于继续训练)
-```
-
-> 注：目前仍在持续标注更多数据，预计最终将有 25,000+ 张标注图片用于训练。
-
-## 项目结构
-
-```
-invoice-master-poc-v2/
-├── src/
-│   ├── cli/              # 命令行工具
-│   │   ├── autolabel.py  # 自动标注
-│   │   ├── train.py      # 模型训练
-│   │   ├── infer.py      # 推理
-│   │   └── serve.py      # Web 服务器
-│   ├── pdf/              # PDF 处理
-│   │   ├── extractor.py  # 文本提取
-│   │   ├── renderer.py   # 图像渲染
-│   │   └── detector.py   # 类型检测
-│   ├── ocr/              # PaddleOCR 封装
-│   │   └── machine_code_parser.py  # 机器可读付款行解析器
-│   ├── normalize/        # 字段规范化
-│   ├── matcher/          # 字段匹配
-│   ├── yolo/             # YOLO 相关
-│   │   ├── annotation_generator.py
-│   │   └── db_dataset.py
-│   ├── inference/        # 推理管道
-│   │   ├── pipeline.py               # 主推理流程
-│   │   ├── yolo_detector.py          # YOLO 检测
-│   │   ├── field_extractor.py        # 字段提取
-│   │   ├── payment_line_parser.py    # 支付行解析器
-│   │   └── customer_number_parser.py # 客户编号解析器
-│   ├── processing/       # 多池处理架构
-│   │   ├── worker_pool.py
-│   │   ├── cpu_pool.py
-│   │   ├── gpu_pool.py
-│   │   ├── task_dispatcher.py
-│   │   └── dual_pool_coordinator.py
-│   ├── web/              # Web 应用
-│   │   ├── app.py        # FastAPI 应用入口
-│   │   ├── routes.py     # API 路由
-│   │   ├── services.py   # 业务逻辑
-│   │   └── schemas.py    # 数据模型
-│   ├── utils/            # 工具模块
-│   │   ├── text_cleaner.py      # 文本清理
-│   │   ├── validators.py        # 字段验证
-│   │   ├── fuzzy_matcher.py     # 模糊匹配
-│   │   └── ocr_corrections.py   # OCR 错误修正
-│   └── data/             # 数据处理
-├── tests/                # 测试文件
-│   ├── ocr/              # OCR 模块测试
-│   │   └── test_machine_code_parser.py
-│   ├── inference/        # 推理模块测试
-│   ├── normalize/        # 规范化模块测试
-│   └── utils/            # 工具模块测试
-├── docs/                 # 文档
-│   ├── REFACTORING_SUMMARY.md
-│   └── TEST_COVERAGE_IMPROVEMENT.md
-├── config.py             # 配置文件
-├── run_server.py         # Web 服务器启动脚本
-├── runs/                 # 训练输出
-│   └── train/
-│       └── invoice_fields/
-│           └── weights/
-│               ├── best.pt      # 最佳模型
-│               └── last.pt      # 最后检查点
-└── requirements.txt
-```
-
-## 多池处理架构
-
-项目使用 CPU + GPU 双池架构处理不同类型的 PDF：
-
-```
-┌─────────────────────────────────────────────────────┐
-│              DualPoolCoordinator                     │
-│  ┌─────────────────┐    ┌─────────────────┐        │
-│  │   CPU Pool      │    │   GPU Pool      │        │
-│  │  (3 workers)    │    │  (1 worker)     │        │
-│  │                 │    │                 │        │
-│  │  Text PDFs      │    │  Scanned PDFs   │        │
-│  │  ~50-87 it/s    │    │  ~1-2 it/s      │        │
-│  └─────────────────┘    └─────────────────┘        │
-│                                                     │
-│  TaskDispatcher: 根据 PDF 类型分配任务               │
-└─────────────────────────────────────────────────────┘
-```
-
-### 关键设计
-
- **spawn 启动方式**: 兼容 CUDA 多进程
- **as_completed()**: 无死锁结果收集
- **进程初始化器**: 每个 worker 加载一次模型
- **协调器持久化**: 跨 CSV 文件复用 worker 池
-
-## 配置文件
-
-### config.py
-
-```python
-# 数据库配置
-DATABASE = {
-    'host': '192.168.68.31',
-    'port': 5432,
-    'database': 'docmaster',
-    'user': 'docmaster',
-    'password': '******',
-}
-
-# 路径配置
-PATHS = {
-    'csv_dir': '~/invoice-data/structured_data',
-    'pdf_dir': '~/invoice-data/raw_pdfs',
-    'output_dir': '~/invoice-data/dataset',
-}
-```
-
-## CLI 命令参考
-
-### autolabel
-
-```bash
-python -m src.cli.autolabel [OPTIONS]
-
-Options:
-  --csv, -c          CSV 文件路径 (支持 glob)
-  --pdf-dir, -p      PDF 文件目录
-  --output, -o       输出目录
-  --workers, -w      单线程模式 worker 数 (默认: 4)
-  --dual-pool        启用双池模式
-  --cpu-workers      CPU 池 worker 数 (默认: 3)
-  --gpu-workers      GPU 池 worker 数 (默认: 1)
-  --dpi              渲染 DPI (默认: 150)
-  --limit, -l        限制处理文档数
-```
-
-### train
-
-```bash
-python -m src.cli.train [OPTIONS]
-
-Options:
-  --model, -m        基础模型路径
-  --epochs, -e       训练轮数 (默认: 100)
-  --batch, -b        批大小 (默认: 16)
-  --imgsz            图像尺寸 (默认: 1280)
-  --dpi              PDF 渲染 DPI (默认: 150)
-  --name             训练名称
-  --limit            限制文档数
-```
-
-### infer
-
-```bash
-python -m src.cli.infer [OPTIONS]
-
-Options:
-  --model, -m        模型路径
-  --input, -i        输入 PDF/图像
-  --output, -o       输出 JSON 路径
-  --confidence       置信度阈值 (默认: 0.5)
-  --dpi              渲染 DPI (默认: 150, 必须与训练DPI一致)
-  --gpu              使用 GPU
-```
-
-### serve
-
-```bash
-python run_server.py [OPTIONS]
-
-Options:
-  --host             绑定地址 (默认: 0.0.0.0)
-  --port             端口 (默认: 8000)
-  --model, -m        模型路径
-  --confidence       置信度阈值 (默认: 0.3)
-  --dpi              渲染 DPI (默认: 150)
-  --no-gpu           禁用 GPU
-  --reload           开发模式自动重载
-  --debug            调试模式
-```
+| 方法 | 端点 | 描述 |
+|------|------|------|
+| POST | `/api/v1/admin/auth/login` | 管理员登录 |
+| GET | `/api/v1/admin/documents` | 文档列表 |
+| POST | `/api/v1/admin/documents/upload` | 上传 PDF |
+| GET | `/api/v1/admin/documents/{id}` | 文档详情 |
+| PATCH | `/api/v1/admin/documents/{id}/status` | 更新文档状态 |
+| POST | `/api/v1/admin/documents/{id}/annotations` | 创建标注 |
+| POST | `/api/v1/admin/training/trigger` | 触发训练任务 |
+| GET | `/api/v1/admin/training/{id}/status` | 查询训练状态 |

 ## Python API

 ```python
-from src.inference.pipeline import InferencePipeline
+from inference.pipeline import InferencePipeline

 # 初始化
 pipeline = InferencePipeline(
@@ -559,41 +291,25 @@ pipeline = InferencePipeline(
 # 处理 PDF
 result = pipeline.process_pdf('invoice.pdf')

-# 处理图片
-result = pipeline.process_image('invoice.png')
-
-# 获取结果
 print(result.fields)
-# {
-#   'InvoiceNumber': '12345',
-#   'Amount': '1234.56',
-#   'payment_line': '# 94228110015950070 # > 48666036#14#',
-#   'customer_number': 'UMJ 436-R',
-#   ...
-# }
+# {'InvoiceNumber': '12345', 'Amount': '1234.56', ...}

-print(result.confidence)   # {'InvoiceNumber': 0.95, 'Amount': 0.92, ...}
-print(result.to_json())    # JSON 格式输出
+print(result.confidence)
+# {'InvoiceNumber': 0.95, 'Amount': 0.92, ...}

-# 访问交叉验证结果
+# 交叉验证
 if result.cross_validation:
    print(f"OCR match: {result.cross_validation.ocr_match}")
-    print(f"Amount match: {result.cross_validation.amount_match}")
-    print(f"Details: {result.cross_validation.details}")
 ```

-### 统一解析器使用
-
 ```python
-from src.inference.payment_line_parser import PaymentLineParser
-from src.inference.customer_number_parser import CustomerNumberParser
+from inference.pipeline.payment_line_parser import PaymentLineParser
+from inference.pipeline.customer_number_parser import CustomerNumberParser

 # Payment Line 解析
 parser = PaymentLineParser()
 result = parser.parse("# 94228110015950070 # 15658 00 8 > 48666036#14#")
-print(f"OCR: {result.ocr_number}")
-print(f"Amount: {result.amount}")
-print(f"Account: {result.account_number}")
+print(f"OCR: {result.ocr_number}, Amount: {result.amount}")

 # Customer Number 解析
 parser = CustomerNumberParser()
@@ -601,156 +317,38 @@ result = parser.parse("Said, Shakar Umj 436-R Billo")
 print(f"Customer Number: {result}")  # "UMJ 436-R"
 ```

+## DPI 配置
+
+系统所有组件统一使用 **150 DPI**。DPI 必须在训练和推理时保持一致。
+
+| 组件 | 配置位置 |
+|------|---------|
+| 全局常量 | `packages/shared/shared/config.py` -> `DEFAULT_DPI = 150` |
+| Web 推理 | `packages/inference/inference/web/config.py` -> `ModelConfig.dpi` |
+| CLI 推理 | `python -m inference.cli.infer --dpi 150` |
+| 自动标注 | `packages/shared/shared/config.py` -> `AUTOLABEL['dpi']` |
+
+## 数据库架构
+
+| 数据库 | 用途 | 存储内容 |
+|--------|------|----------|
+| **PostgreSQL** | 标注结果 | `documents`, `field_results`, `training_tasks` |
+| **SQLite** (AdminDB) | Web 应用 | 文档管理, 标注编辑, 用户认证 |
+
 ## 测试

-### 测试统计
-
-| 指标 | 数值 |
-|------|------|
-| **测试总数** | 688 |
-| **通过率** | 100% |
-| **整体覆盖率** | 37% |
-
-### 关键模块覆盖率
-
-| 模块 | 覆盖率 | 测试数 |
-|------|--------|--------|
-| `machine_code_parser.py` | 65% | 79 |
-| `payment_line_parser.py` | 85% | 45 |
-| `customer_number_parser.py` | 90% | 32 |
-
-### 运行测试
-
 ```bash
 # 运行所有测试
-wsl bash -c "source ~/miniconda3/etc/profile.d/conda.sh && conda activate invoice-py311 && cd /mnt/c/Users/yaoji/git/ColaCoder/invoice-master-poc-v2 && pytest"
+DB_PASSWORD=xxx pytest tests/ -q

 # 运行并查看覆盖率
-wsl bash -c "source ~/miniconda3/etc/profile.d/conda.sh && conda activate invoice-py311 && cd /mnt/c/Users/yaoji/git/ColaCoder/invoice-master-poc-v2 && pytest --cov=src --cov-report=term-missing"
-
-# 运行特定模块测试
-wsl bash -c "source ~/miniconda3/etc/profile.d/conda.sh && conda activate invoice-py311 && cd /mnt/c/Users/yaoji/git/ColaCoder/invoice-master-poc-v2 && pytest tests/ocr/test_machine_code_parser.py -v"
+DB_PASSWORD=xxx pytest tests/ --cov=packages --cov-report=term-missing
 ```

-### 测试结构
-
-```
-tests/
-├── ocr/
-│   ├── test_machine_code_parser.py   # 支付行解析 (79 tests)
-│   └── test_ocr_engine.py            # OCR 引擎测试
-├── inference/
-│   ├── test_payment_line_parser.py   # 支付行解析器
-│   └── test_customer_number_parser.py # 客户编号解析器
-├── normalize/
-│   └── test_normalizers.py           # 字段规范化
-└── utils/
-    └── test_validators.py            # 字段验证
-```
-
-## 开发状态
-
-**已完成功能**:
- [x] 文本层 PDF 自动标注
- [x] 扫描图 OCR 自动标注
- [x] 多策略字段匹配 (精确/子串/规范化)
- [x] PostgreSQL 数据库存储 (断点续传)
- [x] 信号处理和超时保护
- [x] YOLO 训练 (93.5% mAP@0.5, 10 个字段)
- [x] 推理管道
- [x] 字段规范化和验证
- [x] Web 应用 (FastAPI + REST API)
- [x] 增量训练支持
- [x] 内存优化训练 (--low-memory, --resume)
- [x] Payment Line 解析器 (统一模块)
- [x] Customer Number 解析器 (统一模块)
- [x] Payment Line 交叉验证 (OCR, Amount, Account)
- [x] 文档类型识别 (invoice/letter)
- [x] 单元测试覆盖 (688 tests, 37% coverage)
-
-**进行中**:
- [ ] 完成全部 25,000+ 文档标注
- [ ] 多源融合增强 (Multi-source fusion)
- [ ] OCR 错误修正集成
- [ ] 提升测试覆盖率到 60%+
-
-**计划中**:
- [ ] 表格 items 提取
- [ ] 模型量化部署 (ONNX/TensorRT)
- [ ] 多语言支持扩展
-
-## 关键技术特性
-
-### 1. Payment Line 交叉验证
-
-瑞典发票的 payment_line (支付行) 包含完整的支付信息：OCR 参考号、金额、账号。我们实现了交叉验证机制：
-
-```
-Payment Line: # 94228110015950070 # 15658 00 8 > 48666036#14#
-             ↓                     ↓            ↓
-           OCR Number            Amount     Bankgiro Account
-```
-
-**验证流程**:
-1. 从 payment_line 提取 OCR、Amount、Account
-2. 与单独检测的字段对比验证
-3. **payment_line 值优先** - 如有不匹配，采用 payment_line 的值
-4. 返回验证结果和详细信息
-
-**优势**:
- 提高数据准确性 (payment_line 是机器可读格式，更可靠)
- 发现 OCR 或检测错误
- 为数据质量提供信心指标
-
-### 2. 统一解析器架构
-
-采用独立解析器模块处理复杂字段：
-
-**PaymentLineParser**:
- 解析瑞典标准支付行格式
- 提取 OCR、Amount (包含 Kronor + Öre)、Account + Check digits
- 支持多种变体格式
-
-**CustomerNumberParser**:
- 支持多种瑞典客户编号格式 (`UMJ 436-R`, `JTY 576-3`, `FFL 019N`)
- 从混合文本中提取 (如地址行中的客户编号)
- 大小写不敏感，输出统一大写格式
-
-**优势**:
- 代码模块化、可测试
- 易于扩展新格式
- 统一的解析逻辑，减少重复代码
-
-### 3. 文档类型自动识别
-
-根据 payment_line 字段自动判断文档类型：
-
- **invoice**: 包含 payment_line 的发票文档
- **letter**: 不含 payment_line 的信函文档
-
-这个特性帮助下游系统区分处理流程。
-
-### 4. 低内存模式训练
-
-支持在内存受限环境下训练：
-
-```bash
-python -m src.cli.train --low-memory
-```
-
-自动调整:
- batch size: 16 → 8
- workers: 8 → 4
- cache: disabled
- 推荐用于 GPU 内存 < 8GB 或系统内存 < 16GB 的场景
-
-### 5. 断点续传训练
-
-训练中断后可从检查点恢复：
-
-```bash
-python -m src.cli.train --resume --model runs/train/invoice_fields/weights/last.pt
-```
+| 指标 | 数值 |
+|------|------|
+| **测试总数** | 922 |
+| **通过率** | 100% |

 ## 技术栈

@@ -762,32 +360,7 @@ python -m src.cli.train --resume --model runs/train/invoice_fields/weights/last.
 | **数据库** | PostgreSQL + psycopg2 |
 | **Web 框架** | FastAPI + Uvicorn |
 | **深度学习** | PyTorch + CUDA 12.x |
-
-## 常见问题
-
-**Q: 为什么必须在 WSL 环境运行？**
-
-A: PaddleOCR 和某些依赖在 Windows 原生环境存在兼容性问题。WSL 提供完整的 Linux 环境，确保所有依赖正常工作。
-
-**Q: 训练过程中出现 OOM (内存不足) 错误怎么办？**
-
-A: 使用 `--low-memory` 模式，或手动调整 `--batch` 和 `--workers` 参数。
-
-**Q: payment_line 和单独检测字段不匹配时怎么处理？**
-
-A: 系统默认优先采用 payment_line 的值，因为 payment_line 是机器可读格式，通常更准确。验证结果会记录在 `cross_validation` 字段中。
-
-**Q: 如何添加新的字段类型？**
-
-A:
-1. 在 `src/inference/constants.py` 添加字段定义
-2. 在 `field_extractor.py` 添加规范化方法
-3. 重新生成标注数据
-4. 从头训练模型
-
-**Q: 可以用 CPU 训练吗？**
-
-A: 可以，但速度会非常慢 (慢 10-50 倍)。强烈建议使用 GPU 训练。
+| **部署** | Docker + Azure ACI (训练) / App Service (推理) |

 ## 许可证

--- a/docker-compose.yml
+++ b/docker-compose.yml
@@ -0,0 +1,60 @@
+version: "3.8"
+
+services:
+  postgres:
+    image: postgres:15
+    environment:
+      POSTGRES_DB: docmaster
+      POSTGRES_USER: docmaster
+      POSTGRES_PASSWORD: ${DB_PASSWORD:-devpassword}
+    ports:
+      - "5432:5432"
+    volumes:
+      - pgdata:/var/lib/postgresql/data
+      - ./migrations:/docker-entrypoint-initdb.d
+
+  inference:
+    build:
+      context: .
+      dockerfile: packages/inference/Dockerfile
+    ports:
+      - "8000:8000"
+    environment:
+      - DB_HOST=postgres
+      - DB_PORT=5432
+      - DB_NAME=docmaster
+      - DB_USER=docmaster
+      - DB_PASSWORD=${DB_PASSWORD:-devpassword}
+      - MODEL_PATH=/app/models/best.pt
+    volumes:
+      - ./models:/app/models
+    depends_on:
+      - postgres
+
+  training:
+    build:
+      context: .
+      dockerfile: packages/training/Dockerfile
+    environment:
+      - DB_HOST=postgres
+      - DB_PORT=5432
+      - DB_NAME=docmaster
+      - DB_USER=docmaster
+      - DB_PASSWORD=${DB_PASSWORD:-devpassword}
+    volumes:
+      - ./models:/app/models
+      - ./temp:/app/temp
+    depends_on:
+      - postgres
+    # Override CMD for local dev polling mode
+    command: ["python", "run_training.py", "--poll", "--poll-interval", "30"]
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: 1
+              capabilities: [gpu]
+
+volumes:
+  pgdata:
--- a/docs/training-flow.mmd
+++ b/docs/training-flow.mmd
@@ -0,0 +1,54 @@
+flowchart TD
+    A[CLI Entry Point\nsrc/cli/train.py] --> B[Parse Arguments\n--model, --epochs, --batch, --imgsz, etc.]
+    B --> C[Connect PostgreSQL\nDB_HOST / DB_NAME / DB_PASSWORD]
+
+    C --> D[Load Data from DB\nsrc/yolo/db_dataset.py]
+    D --> D1[Scan temp/doc_id/images/\nfor rendered PNGs]
+    D --> D2[Batch load field_results\nfrom database - batch 500]
+
+    D1 --> E[Create DBYOLODataset]
+    D2 --> E
+
+    E --> F[Split Train/Val/Test\n80% / 10% / 10%\nDocument-level, seed=42]
+
+    F --> G[Export to YOLO Format]
+    G --> G1[Copy images to\ntrain/val/test dirs]
+    G --> G2[Generate .txt labels\nclass x_center y_center w h]
+    G --> G3[Generate dataset.yaml\n+ classes.txt]
+    G --> G4[Coordinate Conversion\nPDF points 72DPI -> render DPI\nNormalize to 0-1]
+
+    G1 --> H{--export-only?}
+    G2 --> H
+    G3 --> H
+    G4 --> H
+
+    H -- Yes --> Z[Done - Dataset exported]
+    H -- No --> I[Load YOLO Model]
+
+    I --> I1{--resume?}
+    I1 -- Yes --> I2[Load last.pt checkpoint]
+    I1 -- No --> I3[Load pretrained model\ne.g. yolo11n.pt]
+
+    I2 --> J[Configure Training]
+    I3 --> J
+
+    J --> J1[Conservative Augmentation\nrotation=5 deg, translate=5%\nno flip, no mosaic, no mixup]
+    J --> J2[imgsz=1280, pretrained=True]
+
+    J1 --> K[model.train\nUltralytics Training Loop]
+    J2 --> K
+
+    K --> L[Training Outputs\nruns/train/name/]
+    L --> L1[weights/best.pt\nweights/last.pt]
+    L --> L2[results.csv + results.png\nTraining curves]
+    L --> L3[PR curves, F1 curves\nConfusion matrix]
+
+    L1 --> M[Test Set Validation\nmodel.val split=test]
+    M --> N[Report Metrics\nmAP@0.5 = 93.5%\nmAP@0.5-0.95]
+
+    N --> O[Close DB Connection]
+
+    style A fill:#4a90d9,color:#fff
+    style K fill:#e67e22,color:#fff
+    style N fill:#27ae60,color:#fff
+    style Z fill:#95a5a6,color:#fff
--- a/frontend/package-lock.json
+++ b/frontend/package-lock.json
--- a/frontend/package.json
+++ b/frontend/package.json
@@ -6,27 +6,36 @@
  "scripts": {
    "dev": "vite",
    "build": "vite build",
-    "preview": "vite preview"
+    "preview": "vite preview",
+    "test": "vitest run",
+    "test:watch": "vitest",
+    "test:coverage": "vitest run --coverage"
  },
  "dependencies": {
+    "@tanstack/react-query": "^5.20.0",
+    "axios": "^1.6.7",
+    "clsx": "^2.1.0",
+    "date-fns": "^3.3.0",
+    "lucide-react": "^0.563.0",
    "react": "^19.2.3",
    "react-dom": "^19.2.3",
-    "lucide-react": "^0.563.0",
-    "recharts": "^3.7.0",
-    "axios": "^1.6.7",
    "react-router-dom": "^6.22.0",
-    "zustand": "^4.5.0",
-    "@tanstack/react-query": "^5.20.0",
-    "date-fns": "^3.3.0",
-    "clsx": "^2.1.0"
+    "recharts": "^3.7.0",
+    "zustand": "^4.5.0"
  },
  "devDependencies": {
+    "@testing-library/jest-dom": "^6.9.1",
+    "@testing-library/react": "^16.3.2",
+    "@testing-library/user-event": "^14.6.1",
    "@types/node": "^22.14.0",
    "@vitejs/plugin-react": "^5.0.0",
+    "@vitest/coverage-v8": "^4.0.18",
+    "autoprefixer": "^10.4.17",
+    "jsdom": "^27.4.0",
+    "postcss": "^8.4.35",
+    "tailwindcss": "^3.4.1",
    "typescript": "~5.8.2",
    "vite": "^6.2.0",
-    "tailwindcss": "^3.4.1",
-    "autoprefixer": "^10.4.17",
-    "postcss": "^8.4.35"
+    "vitest": "^4.0.18"
  }
 }
--- a/frontend/src/components/Badge.test.tsx
+++ b/frontend/src/components/Badge.test.tsx
@@ -0,0 +1,32 @@
+import { render, screen } from '@testing-library/react';
+import { describe, it, expect } from 'vitest';
+import { Badge } from './Badge';
+import { DocumentStatus } from '../types';
+
+describe('Badge', () => {
+  it('renders Exported badge with check icon', () => {
+    render(<Badge status="Exported" />);
+    expect(screen.getByText('Exported')).toBeInTheDocument();
+  });
+
+  it('renders Pending status', () => {
+    render(<Badge status={DocumentStatus.PENDING} />);
+    expect(screen.getByText('Pending')).toBeInTheDocument();
+  });
+
+  it('renders Verified status', () => {
+    render(<Badge status={DocumentStatus.VERIFIED} />);
+    expect(screen.getByText('Verified')).toBeInTheDocument();
+  });
+
+  it('renders Labeled status', () => {
+    render(<Badge status={DocumentStatus.LABELED} />);
+    expect(screen.getByText('Labeled')).toBeInTheDocument();
+  });
+
+  it('renders Partial status with warning indicator', () => {
+    render(<Badge status={DocumentStatus.PARTIAL} />);
+    expect(screen.getByText('Partial')).toBeInTheDocument();
+    expect(screen.getByText('!')).toBeInTheDocument();
+  });
+});
--- a/frontend/src/components/Button.test.tsx
+++ b/frontend/src/components/Button.test.tsx
@@ -0,0 +1,38 @@
+import { render, screen } from '@testing-library/react';
+import userEvent from '@testing-library/user-event';
+import { describe, it, expect, vi } from 'vitest';
+import { Button } from './Button';
+
+describe('Button', () => {
+  it('renders children text', () => {
+    render(<Button>Click me</Button>);
+    expect(screen.getByRole('button', { name: 'Click me' })).toBeInTheDocument();
+  });
+
+  it('calls onClick handler', async () => {
+    const user = userEvent.setup();
+    const onClick = vi.fn();
+    render(<Button onClick={onClick}>Click</Button>);
+    await user.click(screen.getByRole('button'));
+    expect(onClick).toHaveBeenCalledOnce();
+  });
+
+  it('is disabled when disabled prop is set', () => {
+    render(<Button disabled>Disabled</Button>);
+    expect(screen.getByRole('button')).toBeDisabled();
+  });
+
+  it('applies variant styles', () => {
+    const { rerender } = render(<Button variant="primary">Primary</Button>);
+    const btn = screen.getByRole('button');
+    expect(btn.className).toContain('bg-warm-text-secondary');
+
+    rerender(<Button variant="secondary">Secondary</Button>);
+    expect(screen.getByRole('button').className).toContain('border');
+  });
+
+  it('applies size styles', () => {
+    render(<Button size="sm">Small</Button>);
+    expect(screen.getByRole('button').className).toContain('h-8');
+  });
+});
--- a/frontend/tests/setup.ts
+++ b/frontend/tests/setup.ts
@@ -0,0 +1 @@
+import '@testing-library/jest-dom';
--- a/frontend/vitest.config.ts
+++ b/frontend/vitest.config.ts
@@ -0,0 +1,19 @@
+/// <reference types="vitest/config" />
+import { defineConfig } from 'vite';
+import react from '@vitejs/plugin-react';
+
+export default defineConfig({
+  plugins: [react()],
+  test: {
+    globals: true,
+    environment: 'jsdom',
+    setupFiles: ['./tests/setup.ts'],
+    include: ['src/**/*.test.{ts,tsx}', 'tests/**/*.test.{ts,tsx}'],
+    coverage: {
+      provider: 'v8',
+      reporter: ['text', 'lcov'],
+      include: ['src/**/*.{ts,tsx}'],
+      exclude: ['src/**/*.test.{ts,tsx}', 'src/main.tsx'],
+    },
+  },
+});
--- a/migrations/003_training_tasks.sql
+++ b/migrations/003_training_tasks.sql
@@ -0,0 +1,18 @@
+-- Training tasks table for async training job management.
+-- Inference service writes pending tasks; training service polls and executes.
+
+CREATE TABLE IF NOT EXISTS training_tasks (
+    task_id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+    status VARCHAR(20) NOT NULL DEFAULT 'pending',
+    config JSONB,
+    created_at TIMESTAMP WITH TIME ZONE NOT NULL DEFAULT NOW(),
+    scheduled_at TIMESTAMP WITH TIME ZONE,
+    started_at TIMESTAMP WITH TIME ZONE,
+    completed_at TIMESTAMP WITH TIME ZONE,
+    error_message TEXT,
+    model_path TEXT,
+    metrics JSONB
+);
+
+CREATE INDEX IF NOT EXISTS idx_training_tasks_status ON training_tasks(status);
+CREATE INDEX IF NOT EXISTS idx_training_tasks_created ON training_tasks(created_at);
--- a/migrations/004_training_datasets.sql
+++ b/migrations/004_training_datasets.sql
@@ -0,0 +1,39 @@
+-- Training Datasets Management
+-- Tracks dataset-document relationships and train/val/test splits
+
+CREATE TABLE IF NOT EXISTS training_datasets (
+    dataset_id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+    name VARCHAR(255) NOT NULL,
+    description TEXT,
+    status VARCHAR(20) NOT NULL DEFAULT 'building',
+    train_ratio FLOAT NOT NULL DEFAULT 0.8,
+    val_ratio FLOAT NOT NULL DEFAULT 0.1,
+    seed INTEGER NOT NULL DEFAULT 42,
+    total_documents INTEGER NOT NULL DEFAULT 0,
+    total_images INTEGER NOT NULL DEFAULT 0,
+    total_annotations INTEGER NOT NULL DEFAULT 0,
+    dataset_path VARCHAR(512),
+    error_message TEXT,
+    created_at TIMESTAMP WITH TIME ZONE NOT NULL DEFAULT NOW(),
+    updated_at TIMESTAMP WITH TIME ZONE NOT NULL DEFAULT NOW()
+);
+
+CREATE INDEX IF NOT EXISTS idx_training_datasets_status ON training_datasets(status);
+
+CREATE TABLE IF NOT EXISTS dataset_documents (
+    id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+    dataset_id UUID NOT NULL REFERENCES training_datasets(dataset_id) ON DELETE CASCADE,
+    document_id UUID NOT NULL REFERENCES admin_documents(document_id),
+    split VARCHAR(10) NOT NULL,
+    page_count INTEGER NOT NULL DEFAULT 0,
+    annotation_count INTEGER NOT NULL DEFAULT 0,
+    created_at TIMESTAMP WITH TIME ZONE NOT NULL DEFAULT NOW(),
+    UNIQUE(dataset_id, document_id)
+);
+
+CREATE INDEX IF NOT EXISTS idx_dataset_documents_dataset ON dataset_documents(dataset_id);
+CREATE INDEX IF NOT EXISTS idx_dataset_documents_document ON dataset_documents(document_id);
+
+-- Add dataset_id to training_tasks
+ALTER TABLE training_tasks ADD COLUMN IF NOT EXISTS dataset_id UUID REFERENCES training_datasets(dataset_id);
+CREATE INDEX IF NOT EXISTS idx_training_tasks_dataset ON training_tasks(dataset_id);
--- a/packages/inference/Dockerfile
+++ b/packages/inference/Dockerfile
@@ -0,0 +1,25 @@
+FROM python:3.11-slim
+
+WORKDIR /app
+
+# Install system dependencies
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    libgl1-mesa-glx libglib2.0-0 libpq-dev gcc \
+    && rm -rf /var/lib/apt/lists/*
+
+# Install shared package
+COPY packages/shared /app/packages/shared
+RUN pip install --no-cache-dir -e /app/packages/shared
+
+# Install inference package
+COPY packages/inference /app/packages/inference
+RUN pip install --no-cache-dir -e /app/packages/inference
+
+# Copy frontend (if needed)
+COPY frontend /app/frontend
+
+WORKDIR /app/packages/inference
+
+EXPOSE 8000
+
+CMD ["python", "run_server.py", "--host", "0.0.0.0", "--port", "8000"]
--- a/packages/inference/inference/init.py
+++ b/packages/inference/inference/init.py
--- a/packages/inference/inference/azure/init.py
+++ b/packages/inference/inference/azure/init.py
--- a/packages/inference/inference/azure/aci_trigger.py
+++ b/packages/inference/inference/azure/aci_trigger.py
@@ -0,0 +1,105 @@
+"""Trigger training jobs on Azure Container Instances."""
+
+import logging
+import os
+
+logger = logging.getLogger(__name__)
+
+# Azure SDK is optional; only needed if using ACI trigger
+try:
+    from azure.identity import DefaultAzureCredential
+    from azure.mgmt.containerinstance import ContainerInstanceManagementClient
+    from azure.mgmt.containerinstance.models import (
+        Container,
+        ContainerGroup,
+        EnvironmentVariable,
+        GpuResource,
+        ResourceRequests,
+        ResourceRequirements,
+    )
+
+    _AZURE_SDK_AVAILABLE = True
+except ImportError:
+    _AZURE_SDK_AVAILABLE = False
+
+
+def start_training_container(task_id: str) -> str | None:
+    """
+    Start an Azure Container Instance for a training task.
+
+    Returns the container group name if successful, None otherwise.
+    Requires environment variables:
+        AZURE_SUBSCRIPTION_ID, AZURE_RESOURCE_GROUP, AZURE_ACR_IMAGE
+    """
+    if not _AZURE_SDK_AVAILABLE:
+        logger.warning(
+            "Azure SDK not installed. Install azure-mgmt-containerinstance "
+            "and azure-identity to use ACI trigger."
+        )
+        return None
+
+    subscription_id = os.environ.get("AZURE_SUBSCRIPTION_ID", "")
+    resource_group = os.environ.get("AZURE_RESOURCE_GROUP", "invoice-training-rg")
+    image = os.environ.get(
+        "AZURE_ACR_IMAGE", "youracr.azurecr.io/invoice-training:latest"
+    )
+    gpu_sku = os.environ.get("AZURE_GPU_SKU", "V100")
+    location = os.environ.get("AZURE_LOCATION", "eastus")
+
+    if not subscription_id:
+        logger.error("AZURE_SUBSCRIPTION_ID not set. Cannot start ACI.")
+        return None
+
+    credential = DefaultAzureCredential()
+    client = ContainerInstanceManagementClient(credential, subscription_id)
+
+    container_name = f"training-{task_id[:8]}"
+
+    env_vars = [
+        EnvironmentVariable(name="TASK_ID", value=task_id),
+    ]
+
+    # Pass DB connection securely
+    for var in ("DB_HOST", "DB_PORT", "DB_NAME", "DB_USER"):
+        val = os.environ.get(var, "")
+        if val:
+            env_vars.append(EnvironmentVariable(name=var, value=val))
+
+    db_password = os.environ.get("DB_PASSWORD", "")
+    if db_password:
+        env_vars.append(
+            EnvironmentVariable(name="DB_PASSWORD", secure_value=db_password)
+        )
+
+    container = Container(
+        name=container_name,
+        image=image,
+        resources=ResourceRequirements(
+            requests=ResourceRequests(
+                cpu=4,
+                memory_in_gb=16,
+                gpu=GpuResource(count=1, sku=gpu_sku),
+            )
+        ),
+        environment_variables=env_vars,
+        command=[
+            "python",
+            "run_training.py",
+            "--task-id",
+            task_id,
+        ],
+    )
+
+    group = ContainerGroup(
+        location=location,
+        containers=[container],
+        os_type="Linux",
+        restart_policy="Never",
+    )
+
+    logger.info("Creating ACI container group: %s", container_name)
+    client.container_groups.begin_create_or_update(
+        resource_group, container_name, group
+    )
+
+    return container_name
--- a/packages/inference/inference/cli/init.py
+++ b/packages/inference/inference/cli/init.py
--- a/packages/inference/inference/cli/infer.py
+++ b/packages/inference/inference/cli/infer.py
@@ -10,8 +10,7 @@ import json
 import sys
 from pathlib import Path

-sys.path.insert(0, str(Path(__file__).parent.parent.parent))
-from src.config import DEFAULT_DPI
+from shared.config import DEFAULT_DPI


 def main():
@@ -91,7 +90,7 @@ def main():
        print(f"Processing {len(pdf_files)} PDF file(s)")
        print(f"Model: {model_path}")

-    from ..inference import InferencePipeline
+    from inference.pipeline import InferencePipeline

    # Initialize pipeline
    pipeline = InferencePipeline(
--- a/packages/inference/inference/cli/serve.py
+++ b/packages/inference/inference/cli/serve.py
@@ -13,9 +13,8 @@ from pathlib import Path

 # Add project root to path
 project_root = Path(__file__).parent.parent.parent
-sys.path.insert(0, str(project_root))

-from src.config import DEFAULT_DPI
+from shared.config import DEFAULT_DPI


 def setup_logging(debug: bool = False) -> None:
@@ -121,7 +120,7 @@ def main() -> None:
    logger.info("=" * 60)

    # Create config
-    from src.web.config import AppConfig, ModelConfig, ServerConfig, StorageConfig
+    from inference.web.config import AppConfig, ModelConfig, ServerConfig, StorageConfig

    config = AppConfig(
        model=ModelConfig(
@@ -142,7 +141,7 @@ def main() -> None:

    # Create and run app
    import uvicorn
-    from src.web.app import create_app
+    from inference.web.app import create_app

    app = create_app(config)

--- a/packages/inference/inference/data/init.py
+++ b/packages/inference/inference/data/init.py
--- a/packages/inference/inference/data/admin_db.py
+++ b/packages/inference/inference/data/admin_db.py
@@ -12,8 +12,8 @@ from uuid import UUID
 from sqlalchemy import func
 from sqlmodel import select

-from src.data.database import get_session_context
-from src.data.admin_models import (
+from inference.data.database import get_session_context
+from inference.data.admin_models import (
    AdminToken,
    AdminDocument,
    AdminAnnotation,
@@ -23,6 +23,8 @@ from src.data.admin_models import (
    BatchUploadFile,
    TrainingDocumentLink,
    AnnotationHistory,
+    TrainingDataset,
+    DatasetDocument,
 )

 logger = logging.getLogger(__name__)
@@ -174,7 +176,7 @@ class AdminDB:

            # For has_annotations filter, we need to join with annotations
            if has_annotations is not None:
-                from src.data.admin_models import AdminAnnotation
+                from inference.data.admin_models import AdminAnnotation

                if has_annotations:
                    # Documents WITH annotations
@@ -200,7 +202,7 @@ class AdminDB:

            # Apply has_annotations filter
            if has_annotations is not None:
-                from src.data.admin_models import AdminAnnotation
+                from inference.data.admin_models import AdminAnnotation

                if has_annotations:
                    statement = (
@@ -456,6 +458,7 @@ class AdminDB:
        scheduled_at: datetime | None = None,
        cron_expression: str | None = None,
        is_recurring: bool = False,
+        dataset_id: str | None = None,
    ) -> str:
        """Create a new training task."""
        with get_session_context() as session:
@@ -469,6 +472,7 @@ class AdminDB:
                cron_expression=cron_expression,
                is_recurring=is_recurring,
                status="scheduled" if scheduled_at else "pending",
+                dataset_id=dataset_id,
            )
            session.add(task)
            session.flush()
@@ -1154,3 +1158,159 @@ class AdminDB:
            session.refresh(annotation)
            session.expunge(annotation)
            return annotation
+
+    # ==========================================================================
+    # Training Dataset Operations
+    # ==========================================================================
+
+    def create_dataset(
+        self,
+        name: str,
+        description: str | None = None,
+        train_ratio: float = 0.8,
+        val_ratio: float = 0.1,
+        seed: int = 42,
+    ) -> TrainingDataset:
+        """Create a new training dataset."""
+        with get_session_context() as session:
+            dataset = TrainingDataset(
+                name=name,
+                description=description,
+                train_ratio=train_ratio,
+                val_ratio=val_ratio,
+                seed=seed,
+            )
+            session.add(dataset)
+            session.commit()
+            session.refresh(dataset)
+            session.expunge(dataset)
+            return dataset
+
+    def get_dataset(self, dataset_id: str | UUID) -> TrainingDataset | None:
+        """Get a dataset by ID."""
+        with get_session_context() as session:
+            dataset = session.get(TrainingDataset, UUID(str(dataset_id)))
+            if dataset:
+                session.expunge(dataset)
+            return dataset
+
+    def get_datasets(
+        self,
+        status: str | None = None,
+        limit: int = 20,
+        offset: int = 0,
+    ) -> tuple[list[TrainingDataset], int]:
+        """List datasets with optional status filter."""
+        with get_session_context() as session:
+            query = select(TrainingDataset)
+            count_query = select(func.count()).select_from(TrainingDataset)
+            if status:
+                query = query.where(TrainingDataset.status == status)
+                count_query = count_query.where(TrainingDataset.status == status)
+            total = session.exec(count_query).one()
+            datasets = session.exec(
+                query.order_by(TrainingDataset.created_at.desc()).offset(offset).limit(limit)
+            ).all()
+            for d in datasets:
+                session.expunge(d)
+            return list(datasets), total
+
+    def update_dataset_status(
+        self,
+        dataset_id: str | UUID,
+        status: str,
+        error_message: str | None = None,
+        total_documents: int | None = None,
+        total_images: int | None = None,
+        total_annotations: int | None = None,
+        dataset_path: str | None = None,
+    ) -> None:
+        """Update dataset status and optional totals."""
+        with get_session_context() as session:
+            dataset = session.get(TrainingDataset, UUID(str(dataset_id)))
+            if not dataset:
+                return
+            dataset.status = status
+            dataset.updated_at = datetime.utcnow()
+            if error_message is not None:
+                dataset.error_message = error_message
+            if total_documents is not None:
+                dataset.total_documents = total_documents
+            if total_images is not None:
+                dataset.total_images = total_images
+            if total_annotations is not None:
+                dataset.total_annotations = total_annotations
+            if dataset_path is not None:
+                dataset.dataset_path = dataset_path
+            session.add(dataset)
+            session.commit()
+
+    def add_dataset_documents(
+        self,
+        dataset_id: str | UUID,
+        documents: list[dict[str, Any]],
+    ) -> None:
+        """Batch insert documents into a dataset.
+
+        Each dict: {document_id, split, page_count, annotation_count}
+        """
+        with get_session_context() as session:
+            for doc in documents:
+                dd = DatasetDocument(
+                    dataset_id=UUID(str(dataset_id)),
+                    document_id=UUID(str(doc["document_id"])),
+                    split=doc["split"],
+                    page_count=doc.get("page_count", 0),
+                    annotation_count=doc.get("annotation_count", 0),
+                )
+                session.add(dd)
+            session.commit()
+
+    def get_dataset_documents(
+        self, dataset_id: str | UUID
+    ) -> list[DatasetDocument]:
+        """Get all documents in a dataset."""
+        with get_session_context() as session:
+            results = session.exec(
+                select(DatasetDocument)
+                .where(DatasetDocument.dataset_id == UUID(str(dataset_id)))
+            ).all()
+            for r in results:
+                session.expunge(r)
+            return list(results)
+
+    def get_documents_by_ids(
+        self, document_ids: list[str]
+    ) -> list[AdminDocument]:
+        """Get documents by list of IDs."""
+        with get_session_context() as session:
+            uuids = [UUID(str(did)) for did in document_ids]
+            results = session.exec(
+                select(AdminDocument).where(AdminDocument.document_id.in_(uuids))
+            ).all()
+            for r in results:
+                session.expunge(r)
+            return list(results)
+
+    def get_annotations_for_document(
+        self, document_id: str | UUID
+    ) -> list[AdminAnnotation]:
+        """Get all annotations for a document."""
+        with get_session_context() as session:
+            results = session.exec(
+                select(AdminAnnotation)
+                .where(AdminAnnotation.document_id == UUID(str(document_id)))
+            ).all()
+            for r in results:
+                session.expunge(r)
+            return list(results)
+
+    def delete_dataset(self, dataset_id: str | UUID) -> bool:
+        """Delete a dataset and its document links (CASCADE)."""
+        with get_session_context() as session:
+            dataset = session.get(TrainingDataset, UUID(str(dataset_id)))
+            if not dataset:
+                return False
+            session.delete(dataset)
+            session.commit()
+            return True
--- a/packages/inference/inference/data/admin_models.py
+++ b/packages/inference/inference/data/admin_models.py
@@ -131,6 +131,7 @@ class TrainingTask(SQLModel, table=True):
    # Status: pending, scheduled, running, completed, failed, cancelled
    task_type: str = Field(default="train", max_length=20)
    # Task type: train, finetune
+    dataset_id: UUID | None = Field(default=None, foreign_key="training_datasets.dataset_id", index=True)
    # Training configuration
    config: dict[str, Any] | None = Field(default=None, sa_column=Column(JSON))
    # Schedule settings
@@ -225,6 +226,42 @@ class BatchUploadFile(SQLModel, table=True):
 # =============================================================================


+class TrainingDataset(SQLModel, table=True):
+    """Training dataset containing selected documents with train/val/test splits."""
+
+    __tablename__ = "training_datasets"
+
+    dataset_id: UUID = Field(default_factory=uuid4, primary_key=True)
+    name: str = Field(max_length=255)
+    description: str | None = Field(default=None)
+    status: str = Field(default="building", max_length=20, index=True)
+    # Status: building, ready, training, archived, failed
+    train_ratio: float = Field(default=0.8)
+    val_ratio: float = Field(default=0.1)
+    seed: int = Field(default=42)
+    total_documents: int = Field(default=0)
+    total_images: int = Field(default=0)
+    total_annotations: int = Field(default=0)
+    dataset_path: str | None = Field(default=None, max_length=512)
+    error_message: str | None = Field(default=None)
+    created_at: datetime = Field(default_factory=datetime.utcnow)
+    updated_at: datetime = Field(default_factory=datetime.utcnow)
+
+
+class DatasetDocument(SQLModel, table=True):
+    """Junction table linking datasets to documents with split assignment."""
+
+    __tablename__ = "dataset_documents"
+
+    id: UUID = Field(default_factory=uuid4, primary_key=True)
+    dataset_id: UUID = Field(foreign_key="training_datasets.dataset_id", index=True)
+    document_id: UUID = Field(foreign_key="admin_documents.document_id", index=True)
+    split: str = Field(max_length=10)  # train, val, test
+    page_count: int = Field(default=0)
+    annotation_count: int = Field(default=0)
+    created_at: datetime = Field(default_factory=datetime.utcnow)
+
+
 class TrainingDocumentLink(SQLModel, table=True):
    """Junction table linking training tasks to documents."""

@@ -336,4 +373,35 @@ class TrainingTaskRead(SQLModel):
    error_message: str | None
    result_metrics: dict[str, Any] | None
    model_path: str | None
+    dataset_id: UUID | None
    created_at: datetime
+
+
+class TrainingDatasetRead(SQLModel):
+    """Training dataset response model."""
+
+    dataset_id: UUID
+    name: str
+    description: str | None
+    status: str
+    train_ratio: float
+    val_ratio: float
+    seed: int
+    total_documents: int
+    total_images: int
+    total_annotations: int
+    dataset_path: str | None
+    error_message: str | None
+    created_at: datetime
+    updated_at: datetime
+
+
+class DatasetDocumentRead(SQLModel):
+    """Dataset document response model."""
+
+    id: UUID
+    dataset_id: UUID
+    document_id: UUID
+    split: str
+    page_count: int
+    annotation_count: int
--- a/packages/inference/inference/data/async_request_db.py
+++ b/packages/inference/inference/data/async_request_db.py
@@ -12,8 +12,8 @@ from uuid import UUID
 from sqlalchemy import func, text
 from sqlmodel import Session, select

-from src.data.database import get_session_context, create_db_and_tables, close_engine
-from src.data.models import ApiKey, AsyncRequest, RateLimitEvent
+from inference.data.database import get_session_context, create_db_and_tables, close_engine
+from inference.data.models import ApiKey, AsyncRequest, RateLimitEvent

 logger = logging.getLogger(__name__)

--- a/packages/inference/inference/data/database.py
+++ b/packages/inference/inference/data/database.py
@@ -13,8 +13,7 @@ from sqlalchemy import text
 from sqlmodel import Session, SQLModel, create_engine

 import sys
-sys.path.insert(0, str(Path(__file__).parent.parent.parent))
-from src.config import get_db_connection_string
+from shared.config import get_db_connection_string

 logger = logging.getLogger(__name__)

@@ -52,8 +51,8 @@ def get_engine():

 def create_db_and_tables() -> None:
    """Create all database tables."""
-    from src.data.models import ApiKey, AsyncRequest, RateLimitEvent  # noqa: F401
-    from src.data.admin_models import (  # noqa: F401
+    from inference.data.models import ApiKey, AsyncRequest, RateLimitEvent  # noqa: F401
+    from inference.data.admin_models import (  # noqa: F401
        AdminToken,
        AdminDocument,
        AdminAnnotation,
--- a/packages/inference/inference/data/models.py
+++ b/packages/inference/inference/data/models.py
--- a/packages/inference/inference/pipeline/init.py
+++ b/packages/inference/inference/pipeline/init.py
--- a/packages/inference/inference/pipeline/constants.py
+++ b/packages/inference/inference/pipeline/constants.py
@@ -92,7 +92,7 @@ constructors or methods. The values here serve as sensible defaults
 based on Swedish invoice processing requirements.

 Example:
-    from src.inference.constants import DEFAULT_CONFIDENCE_THRESHOLD
+    from inference.pipeline.constants import DEFAULT_CONFIDENCE_THRESHOLD

    detector = YOLODetector(
        model_path="model.pt",
--- a/packages/inference/inference/pipeline/customer_number_parser.py
+++ b/packages/inference/inference/pipeline/customer_number_parser.py
@@ -17,7 +17,7 @@ from abc import ABC, abstractmethod
 from dataclasses import dataclass
 from typing import Optional, List

-from src.exceptions import CustomerNumberParseError
+from shared.exceptions import CustomerNumberParseError


@dataclass
--- a/packages/inference/inference/pipeline/field_extractor.py
+++ b/packages/inference/inference/pipeline/field_extractor.py
@@ -4,7 +4,7 @@ Field Extractor Module
 Extracts and validates field values from detected regions.

 This module is used during inference to extract values from OCR text.
-It uses shared utilities from src.utils for text cleaning and validation.
+It uses shared utilities from shared.utils for text cleaning and validation.

 Enhanced features:
 - Multi-source fusion with confidence weighting
@@ -24,10 +24,10 @@ from PIL import Image
 from .yolo_detector import Detection, CLASS_TO_FIELD

 # Import shared utilities for text cleaning and validation
-from src.utils.text_cleaner import TextCleaner
-from src.utils.validators import FieldValidators
-from src.utils.fuzzy_matcher import FuzzyMatcher
-from src.utils.ocr_corrections import OCRCorrections
+from shared.utils.text_cleaner import TextCleaner
+from shared.utils.validators import FieldValidators
+from shared.utils.fuzzy_matcher import FuzzyMatcher
+from shared.utils.ocr_corrections import OCRCorrections

 # Import new unified parsers
 from .payment_line_parser import PaymentLineParser
@@ -104,7 +104,7 @@ class FieldExtractor:
    def ocr_engine(self):
        """Lazy-load OCR engine only when needed."""
        if self._ocr_engine is None:
-            from ..ocr import OCREngine
+            from shared.ocr import OCREngine
            self._ocr_engine = OCREngine(lang=self.ocr_lang)
        return self._ocr_engine

--- a/packages/inference/inference/pipeline/payment_line_parser.py
+++ b/packages/inference/inference/pipeline/payment_line_parser.py
@@ -21,7 +21,7 @@ import logging
 from dataclasses import dataclass
 from typing import Optional

-from src.exceptions import PaymentLineParseError
+from shared.exceptions import PaymentLineParseError


@dataclass
--- a/packages/inference/inference/pipeline/pipeline.py
+++ b/packages/inference/inference/pipeline/pipeline.py
@@ -144,7 +144,7 @@ class InferencePipeline:
        Returns:
            InferenceResult with extracted fields
        """
-        from ..pdf.renderer import render_pdf_to_images
+        from shared.pdf.renderer import render_pdf_to_images
        from PIL import Image
        import io
        import numpy as np
@@ -381,8 +381,8 @@ class InferencePipeline:

    def _run_fallback(self, pdf_path: str | Path, result: InferenceResult) -> None:
        """Run full-page OCR fallback."""
-        from ..pdf.renderer import render_pdf_to_images
-        from ..ocr import OCREngine
+        from shared.pdf.renderer import render_pdf_to_images
+        from shared.ocr import OCREngine
        from PIL import Image
        import io
        import numpy as np
--- a/packages/inference/inference/pipeline/yolo_detector.py
+++ b/packages/inference/inference/pipeline/yolo_detector.py
@@ -189,7 +189,7 @@ class YOLODetector:
        Returns:
            Dict mapping page number to list of detections
        """
-        from ..pdf.renderer import render_pdf_to_images
+        from shared.pdf.renderer import render_pdf_to_images
        from PIL import Image
        import io

--- a/packages/inference/inference/validation/init.py
+++ b/packages/inference/inference/validation/init.py
--- a/packages/inference/inference/validation/llm_validator.py
+++ b/packages/inference/inference/validation/llm_validator.py
@@ -16,7 +16,7 @@ from datetime import datetime
 import psycopg2
 from psycopg2.extras import execute_values

-from src.config import DEFAULT_DPI
+from shared.config import DEFAULT_DPI


@dataclass
--- a/packages/inference/inference/web/init.py
+++ b/packages/inference/inference/web/init.py
--- a/packages/inference/inference/web/admin_routes_new.py
+++ b/packages/inference/inference/web/admin_routes_new.py
@@ -0,0 +1,8 @@
+"""
+Backward compatibility shim for admin_routes.py
+
+DEPRECATED: Import from inference.web.api.v1.admin.documents instead.
+"""
+from inference.web.api.v1.admin.documents import *
+
+__all__ = ["create_admin_router"]
--- a/packages/inference/inference/web/api/init.py
+++ b/packages/inference/inference/web/api/init.py
--- a/packages/inference/inference/web/api/v1/init.py
+++ b/packages/inference/inference/web/api/v1/init.py
--- a/packages/inference/inference/web/api/v1/admin/init.py
+++ b/packages/inference/inference/web/api/v1/admin/init.py
@@ -0,0 +1,19 @@
+"""
+Admin API v1
+
+Document management, annotations, and training endpoints.
+"""
+
+from inference.web.api.v1.admin.annotations import create_annotation_router
+from inference.web.api.v1.admin.auth import create_auth_router
+from inference.web.api.v1.admin.documents import create_documents_router
+from inference.web.api.v1.admin.locks import create_locks_router
+from inference.web.api.v1.admin.training import create_training_router
+
+__all__ = [
+    "create_annotation_router",
+    "create_auth_router",
+    "create_documents_router",
+    "create_locks_router",
+    "create_training_router",
+]
--- a/packages/inference/inference/web/api/v1/admin/annotations.py
+++ b/packages/inference/inference/web/api/v1/admin/annotations.py
@@ -12,11 +12,11 @@ from uuid import UUID
 from fastapi import APIRouter, HTTPException, Query
 from fastapi.responses import FileResponse

-from src.data.admin_db import AdminDB
-from src.data.admin_models import FIELD_CLASSES, FIELD_CLASS_IDS
-from src.web.core.auth import AdminTokenDep, AdminDBDep
-from src.web.services.autolabel import get_auto_label_service
-from src.web.schemas.admin import (
+from inference.data.admin_db import AdminDB
+from inference.data.admin_models import FIELD_CLASSES, FIELD_CLASS_IDS
+from inference.web.core.auth import AdminTokenDep, AdminDBDep
+from inference.web.services.autolabel import get_auto_label_service
+from inference.web.schemas.admin import (
    AnnotationCreate,
    AnnotationItem,
    AnnotationListResponse,
@@ -31,7 +31,7 @@ from src.web.schemas.admin import (
    AutoLabelResponse,
    BoundingBox,
 )
-from src.web.schemas.common import ErrorResponse
+from inference.web.schemas.common import ErrorResponse

 logger = logging.getLogger(__name__)

--- a/packages/inference/inference/web/api/v1/admin/auth.py
+++ b/packages/inference/inference/web/api/v1/admin/auth.py
@@ -10,12 +10,12 @@ from datetime import datetime, timedelta

 from fastapi import APIRouter

-from src.web.core.auth import AdminTokenDep, AdminDBDep
-from src.web.schemas.admin import (
+from inference.web.core.auth import AdminTokenDep, AdminDBDep
+from inference.web.schemas.admin import (
    AdminTokenCreate,
    AdminTokenResponse,
 )
-from src.web.schemas.common import ErrorResponse
+from inference.web.schemas.common import ErrorResponse

 logger = logging.getLogger(__name__)

--- a/packages/inference/inference/web/api/v1/admin/documents.py
+++ b/packages/inference/inference/web/api/v1/admin/documents.py
@@ -11,9 +11,9 @@ from uuid import UUID

 from fastapi import APIRouter, File, HTTPException, Query, UploadFile

-from src.web.config import DEFAULT_DPI, StorageConfig
-from src.web.core.auth import AdminTokenDep, AdminDBDep
-from src.web.schemas.admin import (
+from inference.web.config import DEFAULT_DPI, StorageConfig
+from inference.web.core.auth import AdminTokenDep, AdminDBDep
+from inference.web.schemas.admin import (
    AnnotationItem,
    AnnotationSource,
    AutoLabelStatus,
@@ -27,7 +27,7 @@ from src.web.schemas.admin import (
    ModelMetrics,
    TrainingHistoryItem,
 )
-from src.web.schemas.common import ErrorResponse
+from inference.web.schemas.common import ErrorResponse

 logger = logging.getLogger(__name__)

@@ -142,8 +142,8 @@ def create_documents_router(storage_config: StorageConfig) -> APIRouter:
            raise HTTPException(status_code=500, detail="Failed to save file")

        # Update file path in database
-        from src.data.database import get_session_context
-        from src.data.admin_models import AdminDocument
+        from inference.data.database import get_session_context
+        from inference.data.admin_models import AdminDocument
        with get_session_context() as session:
            doc = session.get(AdminDocument, UUID(document_id))
            if doc:
@@ -520,7 +520,7 @@ def create_documents_router(storage_config: StorageConfig) -> APIRouter:
        # If marking as labeled, save annotations to PostgreSQL DocumentDB
        db_save_result = None
        if status == "labeled":
-            from src.web.services.db_autolabel import save_manual_annotations_to_document_db
+            from inference.web.services.db_autolabel import save_manual_annotations_to_document_db

            # Get all annotations for this document
            annotations = db.get_annotations_for_document(document_id)
--- a/packages/inference/inference/web/api/v1/admin/locks.py
+++ b/packages/inference/inference/web/api/v1/admin/locks.py
@@ -10,12 +10,12 @@ from uuid import UUID

 from fastapi import APIRouter, HTTPException, Query

-from src.web.core.auth import AdminTokenDep, AdminDBDep
-from src.web.schemas.admin import (
+from inference.web.core.auth import AdminTokenDep, AdminDBDep
+from inference.web.schemas.admin import (
    AnnotationLockRequest,
    AnnotationLockResponse,
 )
-from src.web.schemas.common import ErrorResponse
+from inference.web.schemas.common import ErrorResponse

 logger = logging.getLogger(__name__)

--- a/packages/inference/inference/web/api/v1/admin/training/init.py
+++ b/packages/inference/inference/web/api/v1/admin/training/init.py
@@ -0,0 +1,28 @@
+"""
+Admin Training API Routes
+
+FastAPI endpoints for training task management and scheduling.
+"""
+
+from fastapi import APIRouter
+
+from ._utils import _validate_uuid
+from .tasks import register_task_routes
+from .documents import register_document_routes
+from .export import register_export_routes
+from .datasets import register_dataset_routes
+
+
+def create_training_router() -> APIRouter:
+    """Create training API router."""
+    router = APIRouter(prefix="/admin/training", tags=["Admin Training"])
+
+    register_task_routes(router)
+    register_document_routes(router)
+    register_export_routes(router)
+    register_dataset_routes(router)
+
+    return router
+
+
+__all__ = ["create_training_router", "_validate_uuid"]
--- a/packages/inference/inference/web/api/v1/admin/training/_utils.py
+++ b/packages/inference/inference/web/api/v1/admin/training/_utils.py
@@ -0,0 +1,16 @@
+"""Shared utilities for training routes."""
+
+from uuid import UUID
+
+from fastapi import HTTPException
+
+
+def _validate_uuid(value: str, name: str = "ID") -> None:
+    """Validate UUID format."""
+    try:
+        UUID(value)
+    except ValueError:
+        raise HTTPException(
+            status_code=400,
+            detail=f"Invalid {name} format. Must be a valid UUID.",
+        )
--- a/packages/inference/inference/web/api/v1/admin/training/datasets.py
+++ b/packages/inference/inference/web/api/v1/admin/training/datasets.py
@@ -0,0 +1,209 @@
+"""Training Dataset Endpoints."""
+
+import logging
+from typing import Annotated
+
+from fastapi import APIRouter, HTTPException, Query
+
+from inference.web.core.auth import AdminTokenDep, AdminDBDep
+from inference.web.schemas.admin import (
+    DatasetCreateRequest,
+    DatasetDetailResponse,
+    DatasetDocumentItem,
+    DatasetListItem,
+    DatasetListResponse,
+    DatasetResponse,
+    DatasetTrainRequest,
+    TrainingStatus,
+    TrainingTaskResponse,
+)
+
+from ._utils import _validate_uuid
+
+logger = logging.getLogger(__name__)
+
+
+def register_dataset_routes(router: APIRouter) -> None:
+    """Register dataset endpoints on the router."""
+
+    @router.post(
+        "/datasets",
+        response_model=DatasetResponse,
+        summary="Create training dataset",
+        description="Create a dataset from selected documents with train/val/test splits.",
+    )
+    async def create_dataset(
+        request: DatasetCreateRequest,
+        admin_token: AdminTokenDep,
+        db: AdminDBDep,
+    ) -> DatasetResponse:
+        """Create a training dataset from document IDs."""
+        from pathlib import Path
+        from inference.web.services.dataset_builder import DatasetBuilder
+
+        dataset = db.create_dataset(
+            name=request.name,
+            description=request.description,
+            train_ratio=request.train_ratio,
+            val_ratio=request.val_ratio,
+            seed=request.seed,
+        )
+
+        builder = DatasetBuilder(db=db, base_dir=Path("data/datasets"))
+        try:
+            builder.build_dataset(
+                dataset_id=str(dataset.dataset_id),
+                document_ids=request.document_ids,
+                train_ratio=request.train_ratio,
+                val_ratio=request.val_ratio,
+                seed=request.seed,
+                admin_images_dir=Path("data/admin_images"),
+            )
+        except ValueError as e:
+            raise HTTPException(status_code=400, detail=str(e))
+
+        return DatasetResponse(
+            dataset_id=str(dataset.dataset_id),
+            name=dataset.name,
+            status="ready",
+            message="Dataset created successfully",
+        )
+
+    @router.get(
+        "/datasets",
+        response_model=DatasetListResponse,
+        summary="List datasets",
+    )
+    async def list_datasets(
+        admin_token: AdminTokenDep,
+        db: AdminDBDep,
+        status: Annotated[str | None, Query(description="Filter by status")] = None,
+        limit: Annotated[int, Query(ge=1, le=100)] = 20,
+        offset: Annotated[int, Query(ge=0)] = 0,
+    ) -> DatasetListResponse:
+        """List training datasets."""
+        datasets, total = db.get_datasets(status=status, limit=limit, offset=offset)
+        return DatasetListResponse(
+            total=total,
+            limit=limit,
+            offset=offset,
+            datasets=[
+                DatasetListItem(
+                    dataset_id=str(d.dataset_id),
+                    name=d.name,
+                    description=d.description,
+                    status=d.status,
+                    total_documents=d.total_documents,
+                    total_images=d.total_images,
+                    total_annotations=d.total_annotations,
+                    created_at=d.created_at,
+                )
+                for d in datasets
+            ],
+        )
+
+    @router.get(
+        "/datasets/{dataset_id}",
+        response_model=DatasetDetailResponse,
+        summary="Get dataset detail",
+    )
+    async def get_dataset(
+        dataset_id: str,
+        admin_token: AdminTokenDep,
+        db: AdminDBDep,
+    ) -> DatasetDetailResponse:
+        """Get dataset details with document list."""
+        _validate_uuid(dataset_id, "dataset_id")
+        dataset = db.get_dataset(dataset_id)
+        if not dataset:
+            raise HTTPException(status_code=404, detail="Dataset not found")
+
+        docs = db.get_dataset_documents(dataset_id)
+        return DatasetDetailResponse(
+            dataset_id=str(dataset.dataset_id),
+            name=dataset.name,
+            description=dataset.description,
+            status=dataset.status,
+            train_ratio=dataset.train_ratio,
+            val_ratio=dataset.val_ratio,
+            seed=dataset.seed,
+            total_documents=dataset.total_documents,
+            total_images=dataset.total_images,
+            total_annotations=dataset.total_annotations,
+            dataset_path=dataset.dataset_path,
+            error_message=dataset.error_message,
+            documents=[
+                DatasetDocumentItem(
+                    document_id=str(d.document_id),
+                    split=d.split,
+                    page_count=d.page_count,
+                    annotation_count=d.annotation_count,
+                )
+                for d in docs
+            ],
+            created_at=dataset.created_at,
+            updated_at=dataset.updated_at,
+        )
+
+    @router.delete(
+        "/datasets/{dataset_id}",
+        summary="Delete dataset",
+    )
+    async def delete_dataset(
+        dataset_id: str,
+        admin_token: AdminTokenDep,
+        db: AdminDBDep,
+    ) -> dict:
+        """Delete a dataset and its files."""
+        import shutil
+        from pathlib import Path
+
+        _validate_uuid(dataset_id, "dataset_id")
+        dataset = db.get_dataset(dataset_id)
+        if not dataset:
+            raise HTTPException(status_code=404, detail="Dataset not found")
+
+        if dataset.dataset_path:
+            dataset_dir = Path(dataset.dataset_path)
+            if dataset_dir.exists():
+                shutil.rmtree(dataset_dir)
+
+        db.delete_dataset(dataset_id)
+        return {"message": "Dataset deleted"}
+
+    @router.post(
+        "/datasets/{dataset_id}/train",
+        response_model=TrainingTaskResponse,
+        summary="Start training from dataset",
+    )
+    async def train_from_dataset(
+        dataset_id: str,
+        request: DatasetTrainRequest,
+        admin_token: AdminTokenDep,
+        db: AdminDBDep,
+    ) -> TrainingTaskResponse:
+        """Create a training task from a dataset."""
+        _validate_uuid(dataset_id, "dataset_id")
+        dataset = db.get_dataset(dataset_id)
+        if not dataset:
+            raise HTTPException(status_code=404, detail="Dataset not found")
+        if dataset.status != "ready":
+            raise HTTPException(
+                status_code=400,
+                detail=f"Dataset is not ready (status: {dataset.status})",
+            )
+
+        config_dict = request.config.model_dump()
+        task_id = db.create_training_task(
+            admin_token=admin_token,
+            name=request.name,
+            task_type="train",
+            config=config_dict,
+            dataset_id=str(dataset.dataset_id),
+        )
+
+        return TrainingTaskResponse(
+            task_id=task_id,
+            status=TrainingStatus.PENDING,
+            message="Training task created from dataset",
+        )
--- a/packages/inference/inference/web/api/v1/admin/training/documents.py
+++ b/packages/inference/inference/web/api/v1/admin/training/documents.py
@@ -0,0 +1,211 @@
+"""Training Documents and Models Endpoints."""
+
+import logging
+from typing import Annotated
+
+from fastapi import APIRouter, HTTPException, Query
+
+from inference.web.core.auth import AdminTokenDep, AdminDBDep
+from inference.web.schemas.admin import (
+    ModelMetrics,
+    TrainingDocumentItem,
+    TrainingDocumentsResponse,
+    TrainingModelItem,
+    TrainingModelsResponse,
+    TrainingStatus,
+)
+from inference.web.schemas.common import ErrorResponse
+
+from ._utils import _validate_uuid
+
+logger = logging.getLogger(__name__)
+
+
+def register_document_routes(router: APIRouter) -> None:
+    """Register training document and model endpoints on the router."""
+
+    @router.get(
+        "/documents",
+        response_model=TrainingDocumentsResponse,
+        responses={
+            401: {"model": ErrorResponse, "description": "Invalid token"},
+        },
+        summary="Get documents for training",
+        description="Get labeled documents available for training with filtering options.",
+    )
+    async def get_training_documents(
+        admin_token: AdminTokenDep,
+        db: AdminDBDep,
+        has_annotations: Annotated[
+            bool,
+            Query(description="Only include documents with annotations"),
+        ] = True,
+        min_annotation_count: Annotated[
+            int | None,
+            Query(ge=1, description="Minimum annotation count"),
+        ] = None,
+        exclude_used_in_training: Annotated[
+            bool,
+            Query(description="Exclude documents already used in training"),
+        ] = False,
+        limit: Annotated[
+            int,
+            Query(ge=1, le=100, description="Page size"),
+        ] = 100,
+        offset: Annotated[
+            int,
+            Query(ge=0, description="Offset"),
+        ] = 0,
+    ) -> TrainingDocumentsResponse:
+        """Get documents available for training."""
+        documents, total = db.get_documents_for_training(
+            admin_token=admin_token,
+            status="labeled",
+            has_annotations=has_annotations,
+            min_annotation_count=min_annotation_count,
+            exclude_used_in_training=exclude_used_in_training,
+            limit=limit,
+            offset=offset,
+        )
+
+        items = []
+        for doc in documents:
+            annotations = db.get_annotations_for_document(str(doc.document_id))
+
+            sources = {"manual": 0, "auto": 0}
+            for ann in annotations:
+                if ann.source in sources:
+                    sources[ann.source] += 1
+
+            training_links = db.get_document_training_tasks(doc.document_id)
+            used_in_training = [str(link.task_id) for link in training_links]
+
+            items.append(
+                TrainingDocumentItem(
+                    document_id=str(doc.document_id),
+                    filename=doc.filename,
+                    annotation_count=len(annotations),
+                    annotation_sources=sources,
+                    used_in_training=used_in_training,
+                    last_modified=doc.updated_at,
+                )
+            )
+
+        return TrainingDocumentsResponse(
+            total=total,
+            limit=limit,
+            offset=offset,
+            documents=items,
+        )
+
+    @router.get(
+        "/models/{task_id}/download",
+        responses={
+            401: {"model": ErrorResponse, "description": "Invalid token"},
+            404: {"model": ErrorResponse, "description": "Model not found"},
+        },
+        summary="Download trained model",
+        description="Download trained model weights file.",
+    )
+    async def download_model(
+        task_id: str,
+        admin_token: AdminTokenDep,
+        db: AdminDBDep,
+    ):
+        """Download trained model."""
+        from fastapi.responses import FileResponse
+        from pathlib import Path
+
+        _validate_uuid(task_id, "task_id")
+
+        task = db.get_training_task_by_token(task_id, admin_token)
+        if task is None:
+            raise HTTPException(
+                status_code=404,
+                detail="Training task not found or does not belong to this token",
+            )
+
+        if not task.model_path:
+            raise HTTPException(
+                status_code=404,
+                detail="Model file not available for this task",
+            )
+
+        model_path = Path(task.model_path)
+        if not model_path.exists():
+            raise HTTPException(
+                status_code=404,
+                detail="Model file not found on disk",
+            )
+
+        return FileResponse(
+            path=str(model_path),
+            media_type="application/octet-stream",
+            filename=f"{task.name}_model.pt",
+        )
+
+    @router.get(
+        "/models",
+        response_model=TrainingModelsResponse,
+        responses={
+            401: {"model": ErrorResponse, "description": "Invalid token"},
+        },
+        summary="Get trained models",
+        description="Get list of trained models with metrics and download links.",
+    )
+    async def get_training_models(
+        admin_token: AdminTokenDep,
+        db: AdminDBDep,
+        status: Annotated[
+            str | None,
+            Query(description="Filter by status (completed, failed, etc.)"),
+        ] = None,
+        limit: Annotated[
+            int,
+            Query(ge=1, le=100, description="Page size"),
+        ] = 20,
+        offset: Annotated[
+            int,
+            Query(ge=0, description="Offset"),
+        ] = 0,
+    ) -> TrainingModelsResponse:
+        """Get list of trained models."""
+        tasks, total = db.get_training_tasks_by_token(
+            admin_token=admin_token,
+            status=status if status else "completed",
+            limit=limit,
+            offset=offset,
+        )
+
+        items = []
+        for task in tasks:
+            metrics = ModelMetrics(
+                mAP=task.metrics_mAP,
+                precision=task.metrics_precision,
+                recall=task.metrics_recall,
+            )
+
+            download_url = None
+            if task.model_path and task.status == "completed":
+                download_url = f"/api/v1/admin/training/models/{task.task_id}/download"
+
+            items.append(
+                TrainingModelItem(
+                    task_id=str(task.task_id),
+                    name=task.name,
+                    status=TrainingStatus(task.status),
+                    document_count=task.document_count,
+                    created_at=task.created_at,
+                    completed_at=task.completed_at,
+                    metrics=metrics,
+                    model_path=task.model_path,
+                    download_url=download_url,
+                )
+            )
+
+        return TrainingModelsResponse(
+            total=total,
+            limit=limit,
+            offset=offset,
+            models=items,
+        )
--- a/packages/inference/inference/web/api/v1/admin/training/export.py
+++ b/packages/inference/inference/web/api/v1/admin/training/export.py
@@ -0,0 +1,121 @@
+"""Training Export Endpoints."""
+
+import logging
+from datetime import datetime
+
+from fastapi import APIRouter, HTTPException
+
+from inference.web.core.auth import AdminTokenDep, AdminDBDep
+from inference.web.schemas.admin import (
+    ExportRequest,
+    ExportResponse,
+)
+from inference.web.schemas.common import ErrorResponse
+
+logger = logging.getLogger(__name__)
+
+
+def register_export_routes(router: APIRouter) -> None:
+    """Register export endpoints on the router."""
+
+    @router.post(
+        "/export",
+        response_model=ExportResponse,
+        responses={
+            400: {"model": ErrorResponse, "description": "Invalid request"},
+            401: {"model": ErrorResponse, "description": "Invalid token"},
+        },
+        summary="Export annotations",
+        description="Export annotations in YOLO format for training.",
+    )
+    async def export_annotations(
+        request: ExportRequest,
+        admin_token: AdminTokenDep,
+        db: AdminDBDep,
+    ) -> ExportResponse:
+        """Export annotations for training."""
+        from pathlib import Path
+        import shutil
+
+        if request.format not in ("yolo", "coco", "voc"):
+            raise HTTPException(
+                status_code=400,
+                detail=f"Unsupported export format: {request.format}",
+            )
+
+        documents = db.get_labeled_documents_for_export(admin_token)
+
+        if not documents:
+            raise HTTPException(
+                status_code=400,
+                detail="No labeled documents available for export",
+            )
+
+        export_dir = Path("data/exports") / f"export_{datetime.utcnow().strftime('%Y%m%d_%H%M%S')}"
+        export_dir.mkdir(parents=True, exist_ok=True)
+
+        (export_dir / "images" / "train").mkdir(parents=True, exist_ok=True)
+        (export_dir / "images" / "val").mkdir(parents=True, exist_ok=True)
+        (export_dir / "labels" / "train").mkdir(parents=True, exist_ok=True)
+        (export_dir / "labels" / "val").mkdir(parents=True, exist_ok=True)
+
+        total_docs = len(documents)
+        train_count = int(total_docs * request.split_ratio)
+        train_docs = documents[:train_count]
+        val_docs = documents[train_count:]
+
+        total_images = 0
+        total_annotations = 0
+
+        for split, docs in [("train", train_docs), ("val", val_docs)]:
+            for doc in docs:
+                annotations = db.get_annotations_for_document(str(doc.document_id))
+
+                if not annotations:
+                    continue
+
+                for page_num in range(1, doc.page_count + 1):
+                    page_annotations = [a for a in annotations if a.page_number == page_num]
+
+                    if not page_annotations and not request.include_images:
+                        continue
+
+                    src_image = Path("data/admin_images") / str(doc.document_id) / f"page_{page_num}.png"
+                    if not src_image.exists():
+                        continue
+
+                    image_name = f"{doc.document_id}_page{page_num}.png"
+                    dst_image = export_dir / "images" / split / image_name
+                    shutil.copy(src_image, dst_image)
+                    total_images += 1
+
+                    label_name = f"{doc.document_id}_page{page_num}.txt"
+                    label_path = export_dir / "labels" / split / label_name
+
+                    with open(label_path, "w") as f:
+                        for ann in page_annotations:
+                            line = f"{ann.class_id} {ann.x_center:.6f} {ann.y_center:.6f} {ann.width:.6f} {ann.height:.6f}\n"
+                            f.write(line)
+                            total_annotations += 1
+
+        from inference.data.admin_models import FIELD_CLASSES
+
+        yaml_content = f"""# Auto-generated YOLO dataset config
+path: {export_dir.absolute()}
+train: images/train
+val: images/val
+
+nc: {len(FIELD_CLASSES)}
+names: {list(FIELD_CLASSES.values())}
+"""
+        (export_dir / "data.yaml").write_text(yaml_content)
+
+        return ExportResponse(
+            status="completed",
+            export_path=str(export_dir),
+            total_images=total_images,
+            total_annotations=total_annotations,
+            train_count=len(train_docs),
+            val_count=len(val_docs),
+            message=f"Exported {total_images} images with {total_annotations} annotations",
+        )
--- a/packages/inference/inference/web/api/v1/admin/training/tasks.py
+++ b/packages/inference/inference/web/api/v1/admin/training/tasks.py
@@ -0,0 +1,263 @@
+"""Training Task Endpoints."""
+
+import logging
+from typing import Annotated
+
+from fastapi import APIRouter, HTTPException, Query
+
+from inference.web.core.auth import AdminTokenDep, AdminDBDep
+from inference.web.schemas.admin import (
+    TrainingLogItem,
+    TrainingLogsResponse,
+    TrainingStatus,
+    TrainingTaskCreate,
+    TrainingTaskDetailResponse,
+    TrainingTaskItem,
+    TrainingTaskListResponse,
+    TrainingTaskResponse,
+    TrainingType,
+)
+from inference.web.schemas.common import ErrorResponse
+
+from ._utils import _validate_uuid
+
+logger = logging.getLogger(__name__)
+
+
+def register_task_routes(router: APIRouter) -> None:
+    """Register training task endpoints on the router."""
+
+    @router.post(
+        "/tasks",
+        response_model=TrainingTaskResponse,
+        responses={
+            400: {"model": ErrorResponse, "description": "Invalid request"},
+            401: {"model": ErrorResponse, "description": "Invalid token"},
+        },
+        summary="Create training task",
+        description="Create a new training task.",
+    )
+    async def create_training_task(
+        request: TrainingTaskCreate,
+        admin_token: AdminTokenDep,
+        db: AdminDBDep,
+    ) -> TrainingTaskResponse:
+        """Create a new training task."""
+        config_dict = request.config.model_dump() if request.config else {}
+
+        task_id = db.create_training_task(
+            admin_token=admin_token,
+            name=request.name,
+            task_type=request.task_type.value,
+            description=request.description,
+            config=config_dict,
+            scheduled_at=request.scheduled_at,
+            cron_expression=request.cron_expression,
+            is_recurring=bool(request.cron_expression),
+        )
+
+        return TrainingTaskResponse(
+            task_id=task_id,
+            status=TrainingStatus.SCHEDULED if request.scheduled_at else TrainingStatus.PENDING,
+            message="Training task created successfully",
+        )
+
+    @router.get(
+        "/tasks",
+        response_model=TrainingTaskListResponse,
+        responses={
+            401: {"model": ErrorResponse, "description": "Invalid token"},
+        },
+        summary="List training tasks",
+        description="List all training tasks.",
+    )
+    async def list_training_tasks(
+        admin_token: AdminTokenDep,
+        db: AdminDBDep,
+        status: Annotated[
+            str | None,
+            Query(description="Filter by status"),
+        ] = None,
+        limit: Annotated[
+            int,
+            Query(ge=1, le=100, description="Page size"),
+        ] = 20,
+        offset: Annotated[
+            int,
+            Query(ge=0, description="Offset"),
+        ] = 0,
+    ) -> TrainingTaskListResponse:
+        """List training tasks."""
+        valid_statuses = ("pending", "scheduled", "running", "completed", "failed", "cancelled")
+        if status and status not in valid_statuses:
+            raise HTTPException(
+                status_code=400,
+                detail=f"Invalid status: {status}. Must be one of: {', '.join(valid_statuses)}",
+            )
+
+        tasks, total = db.get_training_tasks_by_token(
+            admin_token=admin_token,
+            status=status,
+            limit=limit,
+            offset=offset,
+        )
+
+        items = [
+            TrainingTaskItem(
+                task_id=str(task.task_id),
+                name=task.name,
+                task_type=TrainingType(task.task_type),
+                status=TrainingStatus(task.status),
+                scheduled_at=task.scheduled_at,
+                is_recurring=task.is_recurring,
+                started_at=task.started_at,
+                completed_at=task.completed_at,
+                created_at=task.created_at,
+            )
+            for task in tasks
+        ]
+
+        return TrainingTaskListResponse(
+            total=total,
+            limit=limit,
+            offset=offset,
+            tasks=items,
+        )
+
+    @router.get(
+        "/tasks/{task_id}",
+        response_model=TrainingTaskDetailResponse,
+        responses={
+            401: {"model": ErrorResponse, "description": "Invalid token"},
+            404: {"model": ErrorResponse, "description": "Task not found"},
+        },
+        summary="Get training task detail",
+        description="Get training task details.",
+    )
+    async def get_training_task(
+        task_id: str,
+        admin_token: AdminTokenDep,
+        db: AdminDBDep,
+    ) -> TrainingTaskDetailResponse:
+        """Get training task details."""
+        _validate_uuid(task_id, "task_id")
+
+        task = db.get_training_task_by_token(task_id, admin_token)
+        if task is None:
+            raise HTTPException(
+                status_code=404,
+                detail="Training task not found or does not belong to this token",
+            )
+
+        return TrainingTaskDetailResponse(
+            task_id=str(task.task_id),
+            name=task.name,
+            description=task.description,
+            task_type=TrainingType(task.task_type),
+            status=TrainingStatus(task.status),
+            config=task.config,
+            scheduled_at=task.scheduled_at,
+            cron_expression=task.cron_expression,
+            is_recurring=task.is_recurring,
+            started_at=task.started_at,
+            completed_at=task.completed_at,
+            error_message=task.error_message,
+            result_metrics=task.result_metrics,
+            model_path=task.model_path,
+            created_at=task.created_at,
+        )
+
+    @router.post(
+        "/tasks/{task_id}/cancel",
+        response_model=TrainingTaskResponse,
+        responses={
+            401: {"model": ErrorResponse, "description": "Invalid token"},
+            404: {"model": ErrorResponse, "description": "Task not found"},
+            409: {"model": ErrorResponse, "description": "Cannot cancel task"},
+        },
+        summary="Cancel training task",
+        description="Cancel a pending or scheduled training task.",
+    )
+    async def cancel_training_task(
+        task_id: str,
+        admin_token: AdminTokenDep,
+        db: AdminDBDep,
+    ) -> TrainingTaskResponse:
+        """Cancel a training task."""
+        _validate_uuid(task_id, "task_id")
+
+        task = db.get_training_task_by_token(task_id, admin_token)
+        if task is None:
+            raise HTTPException(
+                status_code=404,
+                detail="Training task not found or does not belong to this token",
+            )
+
+        if task.status not in ("pending", "scheduled"):
+            raise HTTPException(
+                status_code=409,
+                detail=f"Cannot cancel task with status: {task.status}",
+            )
+
+        success = db.cancel_training_task(task_id)
+        if not success:
+            raise HTTPException(
+                status_code=500,
+                detail="Failed to cancel training task",
+            )
+
+        return TrainingTaskResponse(
+            task_id=task_id,
+            status=TrainingStatus.CANCELLED,
+            message="Training task cancelled successfully",
+        )
+
+    @router.get(
+        "/tasks/{task_id}/logs",
+        response_model=TrainingLogsResponse,
+        responses={
+            401: {"model": ErrorResponse, "description": "Invalid token"},
+            404: {"model": ErrorResponse, "description": "Task not found"},
+        },
+        summary="Get training logs",
+        description="Get training task logs.",
+    )
+    async def get_training_logs(
+        task_id: str,
+        admin_token: AdminTokenDep,
+        db: AdminDBDep,
+        limit: Annotated[
+            int,
+            Query(ge=1, le=500, description="Maximum logs to return"),
+        ] = 100,
+        offset: Annotated[
+            int,
+            Query(ge=0, description="Offset"),
+        ] = 0,
+    ) -> TrainingLogsResponse:
+        """Get training logs."""
+        _validate_uuid(task_id, "task_id")
+
+        task = db.get_training_task_by_token(task_id, admin_token)
+        if task is None:
+            raise HTTPException(
+                status_code=404,
+                detail="Training task not found or does not belong to this token",
+            )
+
+        logs = db.get_training_logs(task_id, limit, offset)
+
+        items = [
+            TrainingLogItem(
+                level=log.level,
+                message=log.message,
+                details=log.details,
+                created_at=log.created_at,
+            )
+            for log in logs
+        ]
+
+        return TrainingLogsResponse(
+            task_id=task_id,
+            logs=items,
+        )
--- a/packages/inference/inference/web/api/v1/batch/init.py
+++ b/packages/inference/inference/web/api/v1/batch/init.py
--- a/packages/inference/inference/web/api/v1/batch/routes.py
+++ b/packages/inference/inference/web/api/v1/batch/routes.py
@@ -14,10 +14,10 @@ from uuid import UUID
 from fastapi import APIRouter, Depends, File, HTTPException, UploadFile, Form
 from fastapi.responses import JSONResponse

-from src.data.admin_db import AdminDB
-from src.web.core.auth import validate_admin_token, get_admin_db
-from src.web.services.batch_upload import BatchUploadService, MAX_COMPRESSED_SIZE, MAX_UNCOMPRESSED_SIZE
-from src.web.workers.batch_queue import BatchTask, get_batch_queue
+from inference.data.admin_db import AdminDB
+from inference.web.core.auth import validate_admin_token, get_admin_db
+from inference.web.services.batch_upload import BatchUploadService, MAX_COMPRESSED_SIZE, MAX_UNCOMPRESSED_SIZE
+from inference.web.workers.batch_queue import BatchTask, get_batch_queue

 logger = logging.getLogger(__name__)

--- a/packages/inference/inference/web/api/v1/public/init.py
+++ b/packages/inference/inference/web/api/v1/public/init.py
@@ -0,0 +1,16 @@
+"""
+Public API v1
+
+Customer-facing endpoints for inference, async processing, and labeling.
+"""
+
+from inference.web.api.v1.public.inference import create_inference_router
+from inference.web.api.v1.public.async_api import create_async_router, set_async_service
+from inference.web.api.v1.public.labeling import create_labeling_router
+
+__all__ = [
+    "create_inference_router",
+    "create_async_router",
+    "set_async_service",
+    "create_labeling_router",
+]
--- a/packages/inference/inference/web/api/v1/public/async_api.py
+++ b/packages/inference/inference/web/api/v1/public/async_api.py
@@ -11,13 +11,13 @@ from uuid import UUID

 from fastapi import APIRouter, File, HTTPException, Query, UploadFile

-from src.web.dependencies import (
+from inference.web.dependencies import (
    ApiKeyDep,
    AsyncDBDep,
    PollRateLimitDep,
    SubmitRateLimitDep,
 )
-from src.web.schemas.inference import (
+from inference.web.schemas.inference import (
    AsyncRequestItem,
    AsyncRequestsListResponse,
    AsyncResultResponse,
@@ -27,7 +27,7 @@ from src.web.schemas.inference import (
    DetectionResult,
    InferenceResult,
 )
-from src.web.schemas.common import ErrorResponse
+from inference.web.schemas.common import ErrorResponse


 def _validate_request_id(request_id: str) -> None:
--- a/packages/inference/inference/web/api/v1/public/inference.py
+++ b/packages/inference/inference/web/api/v1/public/inference.py
@@ -15,17 +15,17 @@ from typing import TYPE_CHECKING
 from fastapi import APIRouter, File, HTTPException, UploadFile, status
 from fastapi.responses import FileResponse

-from src.web.schemas.inference import (
+from inference.web.schemas.inference import (
    DetectionResult,
    HealthResponse,
    InferenceResponse,
    InferenceResult,
 )
-from src.web.schemas.common import ErrorResponse
+from inference.web.schemas.common import ErrorResponse

 if TYPE_CHECKING:
-    from src.web.services import InferenceService
-    from src.web.config import StorageConfig
+    from inference.web.services import InferenceService
+    from inference.web.config import StorageConfig

 logger = logging.getLogger(__name__)

--- a/packages/inference/inference/web/api/v1/public/labeling.py
+++ b/packages/inference/inference/web/api/v1/public/labeling.py
@@ -13,13 +13,13 @@ from typing import TYPE_CHECKING

 from fastapi import APIRouter, Depends, File, Form, HTTPException, UploadFile, status

-from src.data.admin_db import AdminDB
-from src.web.schemas.labeling import PreLabelResponse
-from src.web.schemas.common import ErrorResponse
+from inference.data.admin_db import AdminDB
+from inference.web.schemas.labeling import PreLabelResponse
+from inference.web.schemas.common import ErrorResponse

 if TYPE_CHECKING:
-    from src.web.services import InferenceService
-    from src.web.config import StorageConfig
+    from inference.web.services import InferenceService
+    from inference.web.config import StorageConfig

 logger = logging.getLogger(__name__)

--- a/packages/inference/inference/web/app.py
+++ b/packages/inference/inference/web/app.py
@@ -17,10 +17,10 @@ from fastapi.staticfiles import StaticFiles
 from fastapi.responses import HTMLResponse

 from .config import AppConfig, default_config
-from src.web.services import InferenceService
+from inference.web.services import InferenceService

 # Public API imports
-from src.web.api.v1.public import (
+from inference.web.api.v1.public import (
    create_inference_router,
    create_async_router,
    set_async_service,
@@ -28,28 +28,28 @@ from src.web.api.v1.public import (
 )

 # Async processing imports
-from src.data.async_request_db import AsyncRequestDB
-from src.web.workers.async_queue import AsyncTaskQueue
-from src.web.services.async_processing import AsyncProcessingService
-from src.web.dependencies import init_dependencies
-from src.web.core.rate_limiter import RateLimiter
+from inference.data.async_request_db import AsyncRequestDB
+from inference.web.workers.async_queue import AsyncTaskQueue
+from inference.web.services.async_processing import AsyncProcessingService
+from inference.web.dependencies import init_dependencies
+from inference.web.core.rate_limiter import RateLimiter

 # Admin API imports
-from src.web.api.v1.admin import (
+from inference.web.api.v1.admin import (
    create_annotation_router,
    create_auth_router,
    create_documents_router,
    create_locks_router,
    create_training_router,
 )
-from src.web.core.scheduler import start_scheduler, stop_scheduler
-from src.web.core.autolabel_scheduler import start_autolabel_scheduler, stop_autolabel_scheduler
+from inference.web.core.scheduler import start_scheduler, stop_scheduler
+from inference.web.core.autolabel_scheduler import start_autolabel_scheduler, stop_autolabel_scheduler

 # Batch upload imports
-from src.web.api.v1.batch.routes import router as batch_upload_router
-from src.web.workers.batch_queue import init_batch_queue, shutdown_batch_queue
-from src.web.services.batch_upload import BatchUploadService
-from src.data.admin_db import AdminDB
+from inference.web.api.v1.batch.routes import router as batch_upload_router
+from inference.web.workers.batch_queue import init_batch_queue, shutdown_batch_queue
+from inference.web.services.batch_upload import BatchUploadService
+from inference.data.admin_db import AdminDB

 if TYPE_CHECKING:
    from collections.abc import AsyncGenerator
--- a/packages/inference/inference/web/config.py
+++ b/packages/inference/inference/web/config.py
@@ -8,7 +8,7 @@ from dataclasses import dataclass, field
 from pathlib import Path
 from typing import Any

-from src.config import DEFAULT_DPI, PATHS
+from shared.config import DEFAULT_DPI, PATHS


@dataclass(frozen=True)
--- a/packages/inference/inference/web/core/init.py
+++ b/packages/inference/inference/web/core/init.py
@@ -4,10 +4,10 @@ Core Components
 Reusable core functionality: authentication, rate limiting, scheduling.
 """

-from src.web.core.auth import validate_admin_token, get_admin_db, AdminTokenDep, AdminDBDep
-from src.web.core.rate_limiter import RateLimiter
-from src.web.core.scheduler import start_scheduler, stop_scheduler, get_training_scheduler
-from src.web.core.autolabel_scheduler import (
+from inference.web.core.auth import validate_admin_token, get_admin_db, AdminTokenDep, AdminDBDep
+from inference.web.core.rate_limiter import RateLimiter
+from inference.web.core.scheduler import start_scheduler, stop_scheduler, get_training_scheduler
+from inference.web.core.autolabel_scheduler import (
    start_autolabel_scheduler,
    stop_autolabel_scheduler,
    get_autolabel_scheduler,
--- a/packages/inference/inference/web/core/auth.py
+++ b/packages/inference/inference/web/core/auth.py
@@ -9,8 +9,8 @@ from typing import Annotated

 from fastapi import Depends, Header, HTTPException

-from src.data.admin_db import AdminDB
-from src.data.database import get_session_context
+from inference.data.admin_db import AdminDB
+from inference.data.database import get_session_context

 logger = logging.getLogger(__name__)

--- a/packages/inference/inference/web/core/autolabel_scheduler.py
+++ b/packages/inference/inference/web/core/autolabel_scheduler.py
@@ -8,8 +8,8 @@ import logging
 import threading
 from pathlib import Path

-from src.data.admin_db import AdminDB
-from src.web.services.db_autolabel import (
+from inference.data.admin_db import AdminDB
+from inference.web.services.db_autolabel import (
    get_pending_autolabel_documents,
    process_document_autolabel,
 )
--- a/packages/inference/inference/web/core/rate_limiter.py
+++ b/packages/inference/inference/web/core/rate_limiter.py
@@ -13,7 +13,7 @@ from threading import Lock
 from typing import TYPE_CHECKING

 if TYPE_CHECKING:
-    from src.data.async_request_db import AsyncRequestDB
+    from inference.data.async_request_db import AsyncRequestDB

 logger = logging.getLogger(__name__)

--- a/packages/inference/inference/web/core/scheduler.py
+++ b/packages/inference/inference/web/core/scheduler.py
@@ -10,7 +10,7 @@ from datetime import datetime
 from pathlib import Path
 from typing import Any

-from src.data.admin_db import AdminDB
+from inference.data.admin_db import AdminDB

 logger = logging.getLogger(__name__)

@@ -86,7 +86,8 @@ class TrainingScheduler:
                logger.info(f"Starting training task: {task_id}")

                try:
-                    self._execute_task(task_id, task.config or {})
+                    dataset_id = getattr(task, "dataset_id", None)
+                    self._execute_task(task_id, task.config or {}, dataset_id=dataset_id)
                except Exception as e:
                    logger.error(f"Training task {task_id} failed: {e}")
                    self._db.update_training_task_status(
@@ -98,7 +99,9 @@ class TrainingScheduler:
        except Exception as e:
            logger.error(f"Error checking pending tasks: {e}")

-    def _execute_task(self, task_id: str, config: dict[str, Any]) -> None:
+    def _execute_task(
+        self, task_id: str, config: dict[str, Any], dataset_id: str | None = None
+    ) -> None:
        """Execute a training task."""
        # Update status to running
        self._db.update_training_task_status(task_id, "running")
@@ -114,17 +117,25 @@ class TrainingScheduler:
            device = config.get("device", "0")
            project_name = config.get("project_name", "invoice_fields")

-            # Export annotations for training
-            export_result = self._export_training_data(task_id)
-            if not export_result:
-                raise ValueError("Failed to export training data")
-
-            data_yaml = export_result["data_yaml"]
-
-            self._db.add_training_log(
-                task_id, "INFO",
-                f"Exported {export_result['total_images']} images for training",
-            )
+            # Use dataset if available, otherwise export from scratch
+            if dataset_id:
+                dataset = self._db.get_dataset(dataset_id)
+                if not dataset or not dataset.dataset_path:
+                    raise ValueError(f"Dataset {dataset_id} not found or has no path")
+                data_yaml = str(Path(dataset.dataset_path) / "data.yaml")
+                self._db.add_training_log(
+                    task_id, "INFO",
+                    f"Using pre-built dataset: {dataset.name} ({dataset.total_images} images)",
+                )
+            else:
+                export_result = self._export_training_data(task_id)
+                if not export_result:
+                    raise ValueError("Failed to export training data")
+                data_yaml = export_result["data_yaml"]
+                self._db.add_training_log(
+                    task_id, "INFO",
+                    f"Exported {export_result['total_images']} images for training",
+                )

            # Run YOLO training
            result = self._run_yolo_training(
@@ -157,7 +168,7 @@ class TrainingScheduler:
        """Export training data for a task."""
        from pathlib import Path
        import shutil
-        from src.data.admin_models import FIELD_CLASSES
+        from inference.data.admin_models import FIELD_CLASSES

        # Get all labeled documents
        documents = self._db.get_labeled_documents_for_export()
--- a/packages/inference/inference/web/dependencies.py
+++ b/packages/inference/inference/web/dependencies.py
@@ -9,8 +9,8 @@ from typing import Annotated

 from fastapi import Depends, Header, HTTPException, Request

-from src.data.async_request_db import AsyncRequestDB
-from src.web.rate_limiter import RateLimiter
+from inference.data.async_request_db import AsyncRequestDB
+from inference.web.rate_limiter import RateLimiter

 logger = logging.getLogger(__name__)

--- a/packages/inference/inference/web/rate_limiter.py
+++ b/packages/inference/inference/web/rate_limiter.py
@@ -13,7 +13,7 @@ from threading import Lock
 from typing import TYPE_CHECKING

 if TYPE_CHECKING:
-    from src.data.async_request_db import AsyncRequestDB
+    from inference.data.async_request_db import AsyncRequestDB

 logger = logging.getLogger(__name__)

--- a/packages/inference/inference/web/schemas/init.py
+++ b/packages/inference/inference/web/schemas/init.py
@@ -0,0 +1,11 @@
+"""
+API Schemas
+
+Pydantic models for request/response validation.
+"""
+
+# Import everything from sub-modules for backward compatibility
+from inference.web.schemas.common import *  # noqa: F401, F403
+from inference.web.schemas.admin import *  # noqa: F401, F403
+from inference.web.schemas.inference import *  # noqa: F401, F403
+from inference.web.schemas.labeling import *  # noqa: F401, F403
--- a/packages/inference/inference/web/schemas/admin/init.py
+++ b/packages/inference/inference/web/schemas/admin/init.py
@@ -0,0 +1,17 @@
+"""
+Admin API Request/Response Schemas
+
+Pydantic models for admin API validation and serialization.
+"""
+
+from .enums import *  # noqa: F401, F403
+from .auth import *  # noqa: F401, F403
+from .documents import *  # noqa: F401, F403
+from .annotations import *  # noqa: F401, F403
+from .training import *  # noqa: F401, F403
+from .datasets import *  # noqa: F401, F403
+
+# Resolve forward references for DocumentDetailResponse
+from .documents import DocumentDetailResponse
+
+DocumentDetailResponse.model_rebuild()
--- a/packages/inference/inference/web/schemas/admin/annotations.py
+++ b/packages/inference/inference/web/schemas/admin/annotations.py
@@ -0,0 +1,152 @@
+"""Admin Annotation Schemas."""
+
+from datetime import datetime
+
+from pydantic import BaseModel, Field
+
+from .enums import AnnotationSource
+
+
+class BoundingBox(BaseModel):
+    """Bounding box coordinates."""
+
+    x: int = Field(..., ge=0, description="X coordinate (pixels)")
+    y: int = Field(..., ge=0, description="Y coordinate (pixels)")
+    width: int = Field(..., ge=1, description="Width (pixels)")
+    height: int = Field(..., ge=1, description="Height (pixels)")
+
+
+class AnnotationCreate(BaseModel):
+    """Request to create an annotation."""
+
+    page_number: int = Field(default=1, ge=1, description="Page number (1-indexed)")
+    class_id: int = Field(..., ge=0, le=9, description="Class ID (0-9)")
+    bbox: BoundingBox = Field(..., description="Bounding box in pixels")
+    text_value: str | None = Field(None, description="Text value (optional)")
+
+
+class AnnotationUpdate(BaseModel):
+    """Request to update an annotation."""
+
+    class_id: int | None = Field(None, ge=0, le=9, description="New class ID")
+    bbox: BoundingBox | None = Field(None, description="New bounding box")
+    text_value: str | None = Field(None, description="New text value")
+
+
+class AnnotationItem(BaseModel):
+    """Single annotation item."""
+
+    annotation_id: str = Field(..., description="Annotation UUID")
+    page_number: int = Field(..., ge=1, description="Page number")
+    class_id: int = Field(..., ge=0, le=9, description="Class ID")
+    class_name: str = Field(..., description="Class name")
+    bbox: BoundingBox = Field(..., description="Bounding box in pixels")
+    normalized_bbox: dict[str, float] = Field(
+        ..., description="Normalized bbox (x_center, y_center, width, height)"
+    )
+    text_value: str | None = Field(None, description="Text value")
+    confidence: float | None = Field(None, ge=0, le=1, description="Confidence score")
+    source: AnnotationSource = Field(..., description="Annotation source")
+    created_at: datetime = Field(..., description="Creation timestamp")
+
+
+class AnnotationResponse(BaseModel):
+    """Response for annotation operation."""
+
+    annotation_id: str = Field(..., description="Annotation UUID")
+    message: str = Field(..., description="Status message")
+
+
+class AnnotationListResponse(BaseModel):
+    """Response for annotation list."""
+
+    document_id: str = Field(..., description="Document UUID")
+    page_count: int = Field(..., ge=1, description="Total pages")
+    total_annotations: int = Field(..., ge=0, description="Total annotations")
+    annotations: list[AnnotationItem] = Field(
+        default_factory=list, description="Annotation list"
+    )
+
+
+class AnnotationLockRequest(BaseModel):
+    """Request to acquire annotation lock."""
+
+    duration_seconds: int = Field(
+        default=300,
+        ge=60,
+        le=3600,
+        description="Lock duration in seconds (60-3600)",
+    )
+
+
+class AnnotationLockResponse(BaseModel):
+    """Response for annotation lock operation."""
+
+    document_id: str = Field(..., description="Document UUID")
+    locked: bool = Field(..., description="Whether lock was acquired/released")
+    lock_expires_at: datetime | None = Field(
+        None, description="Lock expiration time"
+    )
+    message: str = Field(..., description="Status message")
+
+
+class AutoLabelRequest(BaseModel):
+    """Request to trigger auto-labeling."""
+
+    field_values: dict[str, str] = Field(
+        ...,
+        description="Field values to match (e.g., {'invoice_number': '12345'})",
+    )
+    replace_existing: bool = Field(
+        default=False, description="Replace existing auto annotations"
+    )
+
+
+class AutoLabelResponse(BaseModel):
+    """Response for auto-labeling."""
+
+    document_id: str = Field(..., description="Document UUID")
+    status: str = Field(..., description="Auto-labeling status")
+    annotations_created: int = Field(
+        default=0, ge=0, description="Number of annotations created"
+    )
+    message: str = Field(..., description="Status message")
+
+
+class AnnotationVerifyRequest(BaseModel):
+    """Request to verify an annotation."""
+
+    pass  # No body needed, just POST to verify
+
+
+class AnnotationVerifyResponse(BaseModel):
+    """Response for annotation verification."""
+
+    annotation_id: str = Field(..., description="Annotation UUID")
+    is_verified: bool = Field(..., description="Verification status")
+    verified_at: datetime = Field(..., description="Verification timestamp")
+    verified_by: str = Field(..., description="Admin token who verified")
+    message: str = Field(..., description="Status message")
+
+
+class AnnotationOverrideRequest(BaseModel):
+    """Request to override an annotation."""
+
+    bbox: dict[str, int] | None = Field(
+        None, description="Updated bounding box {x, y, width, height}"
+    )
+    text_value: str | None = Field(None, description="Updated text value")
+    class_id: int | None = Field(None, ge=0, le=9, description="Updated class ID")
+    class_name: str | None = Field(None, description="Updated class name")
+    reason: str | None = Field(None, description="Reason for override")
+
+
+class AnnotationOverrideResponse(BaseModel):
+    """Response for annotation override."""
+
+    annotation_id: str = Field(..., description="Annotation UUID")
+    source: str = Field(..., description="New source (manual)")
+    override_source: str | None = Field(None, description="Original source (auto)")
+    original_annotation_id: str | None = Field(None, description="Original annotation ID")
+    message: str = Field(..., description="Status message")
+    history_id: str = Field(..., description="History record UUID")
--- a/packages/inference/inference/web/schemas/admin/auth.py
+++ b/packages/inference/inference/web/schemas/admin/auth.py
@@ -0,0 +1,23 @@
+"""Admin Auth Schemas."""
+
+from datetime import datetime
+
+from pydantic import BaseModel, Field
+
+
+class AdminTokenCreate(BaseModel):
+    """Request to create an admin token."""
+
+    name: str = Field(..., min_length=1, max_length=255, description="Token name")
+    expires_in_days: int | None = Field(
+        None, ge=1, le=365, description="Token expiration in days (optional)"
+    )
+
+
+class AdminTokenResponse(BaseModel):
+    """Response with created admin token."""
+
+    token: str = Field(..., description="Admin token")
+    name: str = Field(..., description="Token name")
+    expires_at: datetime | None = Field(None, description="Token expiration time")
+    message: str = Field(..., description="Status message")
--- a/packages/inference/inference/web/schemas/admin/datasets.py
+++ b/packages/inference/inference/web/schemas/admin/datasets.py
@@ -0,0 +1,85 @@
+"""Admin Dataset Schemas."""
+
+from datetime import datetime
+
+from pydantic import BaseModel, Field
+
+from .training import TrainingConfig
+
+
+class DatasetCreateRequest(BaseModel):
+    """Request to create a training dataset."""
+
+    name: str = Field(..., min_length=1, max_length=255, description="Dataset name")
+    description: str | None = Field(None, description="Optional description")
+    document_ids: list[str] = Field(..., min_length=1, description="Document UUIDs to include")
+    train_ratio: float = Field(0.8, ge=0.1, le=0.95, description="Training split ratio")
+    val_ratio: float = Field(0.1, ge=0.05, le=0.5, description="Validation split ratio")
+    seed: int = Field(42, description="Random seed for split")
+
+
+class DatasetDocumentItem(BaseModel):
+    """Document within a dataset."""
+
+    document_id: str
+    split: str
+    page_count: int
+    annotation_count: int
+
+
+class DatasetResponse(BaseModel):
+    """Response after creating a dataset."""
+
+    dataset_id: str
+    name: str
+    status: str
+    message: str
+
+
+class DatasetDetailResponse(BaseModel):
+    """Detailed dataset info with documents."""
+
+    dataset_id: str
+    name: str
+    description: str | None
+    status: str
+    train_ratio: float
+    val_ratio: float
+    seed: int
+    total_documents: int
+    total_images: int
+    total_annotations: int
+    dataset_path: str | None
+    error_message: str | None
+    documents: list[DatasetDocumentItem]
+    created_at: datetime
+    updated_at: datetime
+
+
+class DatasetListItem(BaseModel):
+    """Dataset in list view."""
+
+    dataset_id: str
+    name: str
+    description: str | None
+    status: str
+    total_documents: int
+    total_images: int
+    total_annotations: int
+    created_at: datetime
+
+
+class DatasetListResponse(BaseModel):
+    """Paginated dataset list."""
+
+    total: int
+    limit: int
+    offset: int
+    datasets: list[DatasetListItem]
+
+
+class DatasetTrainRequest(BaseModel):
+    """Request to start training from a dataset."""
+
+    name: str = Field(..., min_length=1, max_length=255, description="Training task name")
+    config: TrainingConfig = Field(..., description="Training configuration")
--- a/packages/inference/inference/web/schemas/admin/documents.py
+++ b/packages/inference/inference/web/schemas/admin/documents.py
@@ -0,0 +1,103 @@
+"""Admin Document Schemas."""
+
+from __future__ import annotations
+
+from datetime import datetime
+from typing import TYPE_CHECKING
+
+from pydantic import BaseModel, Field
+
+from .enums import AutoLabelStatus, DocumentStatus
+
+if TYPE_CHECKING:
+    from .annotations import AnnotationItem
+    from .training import TrainingHistoryItem
+
+
+class DocumentUploadResponse(BaseModel):
+    """Response for document upload."""
+
+    document_id: str = Field(..., description="Document UUID")
+    filename: str = Field(..., description="Original filename")
+    file_size: int = Field(..., ge=0, description="File size in bytes")
+    page_count: int = Field(..., ge=1, description="Number of pages")
+    status: DocumentStatus = Field(..., description="Document status")
+    auto_label_started: bool = Field(
+        default=False, description="Whether auto-labeling was started"
+    )
+    message: str = Field(..., description="Status message")
+
+
+class DocumentItem(BaseModel):
+    """Single document in list."""
+
+    document_id: str = Field(..., description="Document UUID")
+    filename: str = Field(..., description="Original filename")
+    file_size: int = Field(..., ge=0, description="File size in bytes")
+    page_count: int = Field(..., ge=1, description="Number of pages")
+    status: DocumentStatus = Field(..., description="Document status")
+    auto_label_status: AutoLabelStatus | None = Field(
+        None, description="Auto-labeling status"
+    )
+    annotation_count: int = Field(default=0, ge=0, description="Number of annotations")
+    upload_source: str = Field(default="ui", description="Upload source (ui or api)")
+    batch_id: str | None = Field(None, description="Batch ID if uploaded via batch")
+    can_annotate: bool = Field(default=True, description="Whether document can be annotated")
+    created_at: datetime = Field(..., description="Creation timestamp")
+    updated_at: datetime = Field(..., description="Last update timestamp")
+
+
+class DocumentListResponse(BaseModel):
+    """Response for document list."""
+
+    total: int = Field(..., ge=0, description="Total documents")
+    limit: int = Field(..., ge=1, description="Page size")
+    offset: int = Field(..., ge=0, description="Current offset")
+    documents: list[DocumentItem] = Field(
+        default_factory=list, description="Document list"
+    )
+
+
+class DocumentDetailResponse(BaseModel):
+    """Response for document detail."""
+
+    document_id: str = Field(..., description="Document UUID")
+    filename: str = Field(..., description="Original filename")
+    file_size: int = Field(..., ge=0, description="File size in bytes")
+    content_type: str = Field(..., description="MIME type")
+    page_count: int = Field(..., ge=1, description="Number of pages")
+    status: DocumentStatus = Field(..., description="Document status")
+    auto_label_status: AutoLabelStatus | None = Field(
+        None, description="Auto-labeling status"
+    )
+    auto_label_error: str | None = Field(None, description="Auto-labeling error")
+    upload_source: str = Field(default="ui", description="Upload source (ui or api)")
+    batch_id: str | None = Field(None, description="Batch ID if uploaded via batch")
+    csv_field_values: dict[str, str] | None = Field(
+        None, description="CSV field values if uploaded via batch"
+    )
+    can_annotate: bool = Field(default=True, description="Whether document can be annotated")
+    annotation_lock_until: datetime | None = Field(
+        None, description="Lock expiration time if document is locked"
+    )
+    annotations: list["AnnotationItem"] = Field(
+        default_factory=list, description="Document annotations"
+    )
+    image_urls: list[str] = Field(
+        default_factory=list, description="URLs to page images"
+    )
+    training_history: list["TrainingHistoryItem"] = Field(
+        default_factory=list, description="Training tasks that used this document"
+    )
+    created_at: datetime = Field(..., description="Creation timestamp")
+    updated_at: datetime = Field(..., description="Last update timestamp")
+
+
+class DocumentStatsResponse(BaseModel):
+    """Document statistics response."""
+
+    total: int = Field(..., ge=0, description="Total documents")
+    pending: int = Field(default=0, ge=0, description="Pending documents")
+    auto_labeling: int = Field(default=0, ge=0, description="Auto-labeling documents")
+    labeled: int = Field(default=0, ge=0, description="Labeled documents")
+    exported: int = Field(default=0, ge=0, description="Exported documents")
--- a/packages/inference/inference/web/schemas/admin/enums.py
+++ b/packages/inference/inference/web/schemas/admin/enums.py
@@ -0,0 +1,46 @@
+"""Admin API Enums."""
+
+from enum import Enum
+
+
+class DocumentStatus(str, Enum):
+    """Document status enum."""
+
+    PENDING = "pending"
+    AUTO_LABELING = "auto_labeling"
+    LABELED = "labeled"
+    EXPORTED = "exported"
+
+
+class AutoLabelStatus(str, Enum):
+    """Auto-labeling status enum."""
+
+    RUNNING = "running"
+    COMPLETED = "completed"
+    FAILED = "failed"
+
+
+class TrainingStatus(str, Enum):
+    """Training task status enum."""
+
+    PENDING = "pending"
+    SCHEDULED = "scheduled"
+    RUNNING = "running"
+    COMPLETED = "completed"
+    FAILED = "failed"
+    CANCELLED = "cancelled"
+
+
+class TrainingType(str, Enum):
+    """Training task type enum."""
+
+    TRAIN = "train"
+    FINETUNE = "finetune"
+
+
+class AnnotationSource(str, Enum):
+    """Annotation source enum."""
+
+    MANUAL = "manual"
+    AUTO = "auto"
+    IMPORTED = "imported"
--- a/packages/inference/inference/web/schemas/admin/training.py
+++ b/packages/inference/inference/web/schemas/admin/training.py
@@ -0,0 +1,202 @@
+"""Admin Training Schemas."""
+
+from datetime import datetime
+from typing import Any
+
+from pydantic import BaseModel, Field
+
+from .enums import TrainingStatus, TrainingType
+
+
+class TrainingConfig(BaseModel):
+    """Training configuration."""
+
+    model_name: str = Field(default="yolo11n.pt", description="Base model name")
+    epochs: int = Field(default=100, ge=1, le=1000, description="Training epochs")
+    batch_size: int = Field(default=16, ge=1, le=128, description="Batch size")
+    image_size: int = Field(default=640, ge=320, le=1280, description="Image size")
+    learning_rate: float = Field(default=0.01, gt=0, le=1, description="Learning rate")
+    device: str = Field(default="0", description="Device (0 for GPU, cpu for CPU)")
+    project_name: str = Field(
+        default="invoice_fields", description="Training project name"
+    )
+
+
+class TrainingTaskCreate(BaseModel):
+    """Request to create a training task."""
+
+    name: str = Field(..., min_length=1, max_length=255, description="Task name")
+    description: str | None = Field(None, max_length=1000, description="Description")
+    task_type: TrainingType = Field(
+        default=TrainingType.TRAIN, description="Task type"
+    )
+    config: TrainingConfig = Field(
+        default_factory=TrainingConfig, description="Training configuration"
+    )
+    scheduled_at: datetime | None = Field(
+        None, description="Scheduled execution time"
+    )
+    cron_expression: str | None = Field(
+        None, max_length=50, description="Cron expression for recurring tasks"
+    )
+
+
+class TrainingTaskItem(BaseModel):
+    """Single training task in list."""
+
+    task_id: str = Field(..., description="Task UUID")
+    name: str = Field(..., description="Task name")
+    task_type: TrainingType = Field(..., description="Task type")
+    status: TrainingStatus = Field(..., description="Task status")
+    scheduled_at: datetime | None = Field(None, description="Scheduled time")
+    is_recurring: bool = Field(default=False, description="Is recurring task")
+    started_at: datetime | None = Field(None, description="Start time")
+    completed_at: datetime | None = Field(None, description="Completion time")
+    created_at: datetime = Field(..., description="Creation timestamp")
+
+
+class TrainingTaskListResponse(BaseModel):
+    """Response for training task list."""
+
+    total: int = Field(..., ge=0, description="Total tasks")
+    limit: int = Field(..., ge=1, description="Page size")
+    offset: int = Field(..., ge=0, description="Current offset")
+    tasks: list[TrainingTaskItem] = Field(default_factory=list, description="Task list")
+
+
+class TrainingTaskDetailResponse(BaseModel):
+    """Response for training task detail."""
+
+    task_id: str = Field(..., description="Task UUID")
+    name: str = Field(..., description="Task name")
+    description: str | None = Field(None, description="Description")
+    task_type: TrainingType = Field(..., description="Task type")
+    status: TrainingStatus = Field(..., description="Task status")
+    config: dict[str, Any] | None = Field(None, description="Training configuration")
+    scheduled_at: datetime | None = Field(None, description="Scheduled time")
+    cron_expression: str | None = Field(None, description="Cron expression")
+    is_recurring: bool = Field(default=False, description="Is recurring task")
+    started_at: datetime | None = Field(None, description="Start time")
+    completed_at: datetime | None = Field(None, description="Completion time")
+    error_message: str | None = Field(None, description="Error message")
+    result_metrics: dict[str, Any] | None = Field(None, description="Result metrics")
+    model_path: str | None = Field(None, description="Trained model path")
+    created_at: datetime = Field(..., description="Creation timestamp")
+
+
+class TrainingTaskResponse(BaseModel):
+    """Response for training task operation."""
+
+    task_id: str = Field(..., description="Task UUID")
+    status: TrainingStatus = Field(..., description="Task status")
+    message: str = Field(..., description="Status message")
+
+
+class TrainingLogItem(BaseModel):
+    """Single training log entry."""
+
+    level: str = Field(..., description="Log level")
+    message: str = Field(..., description="Log message")
+    details: dict[str, Any] | None = Field(None, description="Additional details")
+    created_at: datetime = Field(..., description="Timestamp")
+
+
+class TrainingLogsResponse(BaseModel):
+    """Response for training logs."""
+
+    task_id: str = Field(..., description="Task UUID")
+    logs: list[TrainingLogItem] = Field(default_factory=list, description="Log entries")
+
+
+class ExportRequest(BaseModel):
+    """Request to export annotations."""
+
+    format: str = Field(
+        default="yolo", description="Export format (yolo, coco, voc)"
+    )
+    include_images: bool = Field(
+        default=True, description="Include images in export"
+    )
+    split_ratio: float = Field(
+        default=0.8, ge=0.5, le=1.0, description="Train/val split ratio"
+    )
+
+
+class ExportResponse(BaseModel):
+    """Response for export operation."""
+
+    status: str = Field(..., description="Export status")
+    export_path: str = Field(..., description="Path to exported dataset")
+    total_images: int = Field(..., ge=0, description="Total images exported")
+    total_annotations: int = Field(..., ge=0, description="Total annotations")
+    train_count: int = Field(..., ge=0, description="Training set count")
+    val_count: int = Field(..., ge=0, description="Validation set count")
+    message: str = Field(..., description="Status message")
+
+
+class TrainingDocumentItem(BaseModel):
+    """Document item for training page."""
+
+    document_id: str = Field(..., description="Document UUID")
+    filename: str = Field(..., description="Filename")
+    annotation_count: int = Field(..., ge=0, description="Total annotations")
+    annotation_sources: dict[str, int] = Field(
+        ..., description="Annotation counts by source (manual, auto)"
+    )
+    used_in_training: list[str] = Field(
+        default_factory=list, description="List of training task IDs that used this document"
+    )
+    last_modified: datetime = Field(..., description="Last modification time")
+
+
+class TrainingDocumentsResponse(BaseModel):
+    """Response for GET /admin/training/documents."""
+
+    total: int = Field(..., ge=0, description="Total document count")
+    limit: int = Field(..., ge=1, le=100, description="Page size")
+    offset: int = Field(..., ge=0, description="Pagination offset")
+    documents: list[TrainingDocumentItem] = Field(
+        default_factory=list, description="Documents available for training"
+    )
+
+
+class ModelMetrics(BaseModel):
+    """Training model metrics."""
+
+    mAP: float | None = Field(None, ge=0.0, le=1.0, description="Mean Average Precision")
+    precision: float | None = Field(None, ge=0.0, le=1.0, description="Precision")
+    recall: float | None = Field(None, ge=0.0, le=1.0, description="Recall")
+
+
+class TrainingModelItem(BaseModel):
+    """Trained model item for model list."""
+
+    task_id: str = Field(..., description="Training task UUID")
+    name: str = Field(..., description="Model name")
+    status: TrainingStatus = Field(..., description="Training status")
+    document_count: int = Field(..., ge=0, description="Documents used in training")
+    created_at: datetime = Field(..., description="Creation timestamp")
+    completed_at: datetime | None = Field(None, description="Completion timestamp")
+    metrics: ModelMetrics = Field(..., description="Model metrics")
+    model_path: str | None = Field(None, description="Path to model weights")
+    download_url: str | None = Field(None, description="Download URL for model")
+
+
+class TrainingModelsResponse(BaseModel):
+    """Response for GET /admin/training/models."""
+
+    total: int = Field(..., ge=0, description="Total model count")
+    limit: int = Field(..., ge=1, le=100, description="Page size")
+    offset: int = Field(..., ge=0, description="Pagination offset")
+    models: list[TrainingModelItem] = Field(
+        default_factory=list, description="Trained models"
+    )
+
+
+class TrainingHistoryItem(BaseModel):
+    """Training history for a document."""
+
+    task_id: str = Field(..., description="Training task UUID")
+    name: str = Field(..., description="Training task name")
+    trained_at: datetime = Field(..., description="Training timestamp")
+    model_metrics: ModelMetrics | None = Field(None, description="Model metrics")
--- a/packages/inference/inference/web/schemas/common.py
+++ b/packages/inference/inference/web/schemas/common.py
--- a/packages/inference/inference/web/schemas/inference.py
+++ b/packages/inference/inference/web/schemas/inference.py
--- a/packages/inference/inference/web/schemas/labeling.py
+++ b/packages/inference/inference/web/schemas/labeling.py
--- a/packages/inference/inference/web/services/init.py
+++ b/packages/inference/inference/web/services/init.py
@@ -0,0 +1,18 @@
+"""
+Business Logic Services
+
+Service layer for processing requests and orchestrating data operations.
+"""
+
+from inference.web.services.autolabel import AutoLabelService, get_auto_label_service
+from inference.web.services.inference import InferenceService
+from inference.web.services.async_processing import AsyncProcessingService
+from inference.web.services.batch_upload import BatchUploadService
+
+__all__ = [
+    "AutoLabelService",
+    "get_auto_label_service",
+    "InferenceService",
+    "AsyncProcessingService",
+    "BatchUploadService",
+]
--- a/packages/inference/inference/web/services/async_processing.py
+++ b/packages/inference/inference/web/services/async_processing.py
@@ -14,13 +14,13 @@ from pathlib import Path
 from threading import Event, Thread
 from typing import TYPE_CHECKING

-from src.data.async_request_db import AsyncRequestDB
-from src.web.workers.async_queue import AsyncTask, AsyncTaskQueue
-from src.web.core.rate_limiter import RateLimiter
+from inference.data.async_request_db import AsyncRequestDB
+from inference.web.workers.async_queue import AsyncTask, AsyncTaskQueue
+from inference.web.core.rate_limiter import RateLimiter

 if TYPE_CHECKING:
-    from src.web.config import AsyncConfig, StorageConfig
-    from src.web.services.inference import InferenceService
+    from inference.web.config import AsyncConfig, StorageConfig
+    from inference.web.services.inference import InferenceService

 logger = logging.getLogger(__name__)

--- a/packages/inference/inference/web/services/autolabel.py
+++ b/packages/inference/inference/web/services/autolabel.py
@@ -11,11 +11,11 @@ from typing import Any
 import numpy as np
 from PIL import Image

-from src.config import DEFAULT_DPI
-from src.data.admin_db import AdminDB
-from src.data.admin_models import FIELD_CLASS_IDS, FIELD_CLASSES
-from src.matcher.field_matcher import FieldMatcher
-from src.ocr.paddle_ocr import OCREngine, OCRToken
+from shared.config import DEFAULT_DPI
+from inference.data.admin_db import AdminDB
+from inference.data.admin_models import FIELD_CLASS_IDS, FIELD_CLASSES
+from shared.matcher.field_matcher import FieldMatcher
+from shared.ocr.paddle_ocr import OCREngine, OCRToken

 logger = logging.getLogger(__name__)

@@ -144,7 +144,7 @@ class AutoLabelService:
        db: AdminDB,
    ) -> int:
        """Process PDF document and create annotations."""
-        from src.pdf.renderer import render_pdf_to_images
+        from shared.pdf.renderer import render_pdf_to_images
        import io

        total_annotations = 0
@@ -222,7 +222,7 @@ class AutoLabelService:
        image_height: int,
    ) -> list[dict[str, Any]]:
        """Find annotations for field values using token matching."""
-        from src.normalize import normalize_field
+        from shared.normalize import normalize_field

        annotations = []

--- a/packages/inference/inference/web/services/batch_upload.py
+++ b/packages/inference/inference/web/services/batch_upload.py
@@ -15,8 +15,8 @@ from uuid import UUID

 from pydantic import BaseModel, Field, field_validator

-from src.data.admin_db import AdminDB
-from src.data.admin_models import CSV_TO_CLASS_MAPPING
+from inference.data.admin_db import AdminDB
+from inference.data.admin_models import CSV_TO_CLASS_MAPPING

 logger = logging.getLogger(__name__)

--- a/packages/inference/inference/web/services/dataset_builder.py
+++ b/packages/inference/inference/web/services/dataset_builder.py
@@ -0,0 +1,188 @@
+"""
+Dataset Builder Service
+
+Creates training datasets by copying images from admin storage,
+generating YOLO label files, and splitting into train/val/test sets.
+"""
+
+import logging
+import random
+import shutil
+from pathlib import Path
+
+import yaml
+
+from inference.data.admin_models import FIELD_CLASSES
+
+logger = logging.getLogger(__name__)
+
+
+class DatasetBuilder:
+    """Builds YOLO training datasets from admin documents."""
+
+    def __init__(self, db, base_dir: Path):
+        self._db = db
+        self._base_dir = Path(base_dir)
+
+    def build_dataset(
+        self,
+        dataset_id: str,
+        document_ids: list[str],
+        train_ratio: float,
+        val_ratio: float,
+        seed: int,
+        admin_images_dir: Path,
+    ) -> dict:
+        """Build a complete YOLO dataset from document IDs.
+
+        Args:
+            dataset_id: UUID of the dataset record.
+            document_ids: List of document UUIDs to include.
+            train_ratio: Fraction for training set.
+            val_ratio: Fraction for validation set.
+            seed: Random seed for reproducible splits.
+            admin_images_dir: Root directory of admin images.
+
+        Returns:
+            Summary dict with total_documents, total_images, total_annotations.
+
+        Raises:
+            ValueError: If no valid documents found.
+        """
+        try:
+            return self._do_build(
+                dataset_id, document_ids, train_ratio, val_ratio, seed, admin_images_dir
+            )
+        except Exception as e:
+            self._db.update_dataset_status(
+                dataset_id=dataset_id,
+                status="failed",
+                error_message=str(e),
+            )
+            raise
+
+    def _do_build(
+        self,
+        dataset_id: str,
+        document_ids: list[str],
+        train_ratio: float,
+        val_ratio: float,
+        seed: int,
+        admin_images_dir: Path,
+    ) -> dict:
+        # 1. Fetch documents
+        documents = self._db.get_documents_by_ids(document_ids)
+        if not documents:
+            raise ValueError("No valid documents found for the given IDs")
+
+        # 2. Create directory structure
+        dataset_dir = self._base_dir / dataset_id
+        for split in ["train", "val", "test"]:
+            (dataset_dir / "images" / split).mkdir(parents=True, exist_ok=True)
+            (dataset_dir / "labels" / split).mkdir(parents=True, exist_ok=True)
+
+        # 3. Shuffle and split documents
+        doc_list = list(documents)
+        rng = random.Random(seed)
+        rng.shuffle(doc_list)
+
+        n = len(doc_list)
+        n_train = max(1, round(n * train_ratio))
+        n_val = max(0, round(n * val_ratio))
+        n_test = n - n_train - n_val
+
+        splits = (
+            ["train"] * n_train
+            + ["val"] * n_val
+            + ["test"] * n_test
+        )
+
+        # 4. Process each document
+        total_images = 0
+        total_annotations = 0
+        dataset_docs = []
+
+        for doc, split in zip(doc_list, splits):
+            doc_id = str(doc.document_id)
+            annotations = self._db.get_annotations_for_document(doc.document_id)
+
+            # Group annotations by page
+            page_annotations: dict[int, list] = {}
+            for ann in annotations:
+                page_annotations.setdefault(ann.page_number, []).append(ann)
+
+            doc_image_count = 0
+            doc_ann_count = 0
+
+            # Copy images and write labels for each page
+            for page_num in range(1, doc.page_count + 1):
+                src_image = Path(admin_images_dir) / doc_id / f"page_{page_num}.png"
+                if not src_image.exists():
+                    logger.warning("Image not found: %s", src_image)
+                    continue
+
+                dst_name = f"{doc_id}_page{page_num}"
+                dst_image = dataset_dir / "images" / split / f"{dst_name}.png"
+                shutil.copy2(src_image, dst_image)
+                doc_image_count += 1
+
+                # Write YOLO label file
+                page_anns = page_annotations.get(page_num, [])
+                label_lines = []
+                for ann in page_anns:
+                    label_lines.append(
+                        f"{ann.class_id} {ann.x_center:.6f} {ann.y_center:.6f} "
+                        f"{ann.width:.6f} {ann.height:.6f}"
+                    )
+                    doc_ann_count += 1
+
+                label_path = dataset_dir / "labels" / split / f"{dst_name}.txt"
+                label_path.write_text("\n".join(label_lines))
+
+            total_images += doc_image_count
+            total_annotations += doc_ann_count
+
+            dataset_docs.append({
+                "document_id": doc_id,
+                "split": split,
+                "page_count": doc_image_count,
+                "annotation_count": doc_ann_count,
+            })
+
+        # 5. Record document-split assignments in DB
+        self._db.add_dataset_documents(
+            dataset_id=dataset_id,
+            documents=dataset_docs,
+        )
+
+        # 6. Generate data.yaml
+        self._generate_data_yaml(dataset_dir)
+
+        # 7. Update dataset status
+        self._db.update_dataset_status(
+            dataset_id=dataset_id,
+            status="ready",
+            total_documents=len(doc_list),
+            total_images=total_images,
+            total_annotations=total_annotations,
+            dataset_path=str(dataset_dir),
+        )
+
+        return {
+            "total_documents": len(doc_list),
+            "total_images": total_images,
+            "total_annotations": total_annotations,
+        }
+
+    def _generate_data_yaml(self, dataset_dir: Path) -> None:
+        """Generate YOLO data.yaml configuration file."""
+        data = {
+            "path": str(dataset_dir.absolute()),
+            "train": "images/train",
+            "val": "images/val",
+            "test": "images/test",
+            "nc": len(FIELD_CLASSES),
+            "names": FIELD_CLASSES,
+        }
+        yaml_path = dataset_dir / "data.yaml"
+        yaml_path.write_text(yaml.dump(data, default_flow_style=False, allow_unicode=True))
--- a/packages/inference/inference/web/services/db_autolabel.py
+++ b/packages/inference/inference/web/services/db_autolabel.py
@@ -11,11 +11,11 @@ import logging
 from pathlib import Path
 from typing import Any

-from src.config import DEFAULT_DPI
-from src.data.admin_db import AdminDB
-from src.data.admin_models import AdminDocument, CSV_TO_CLASS_MAPPING
-from src.data.db import DocumentDB
-from src.web.config import StorageConfig
+from shared.config import DEFAULT_DPI
+from inference.data.admin_db import AdminDB
+from inference.data.admin_models import AdminDocument, CSV_TO_CLASS_MAPPING
+from shared.data.db import DocumentDB
+from inference.web.config import StorageConfig

 logger = logging.getLogger(__name__)

@@ -81,8 +81,8 @@ def get_pending_autolabel_documents(
        List of AdminDocument records with status='auto_labeling' and auto_label_status='pending'
    """
    from sqlmodel import select
-    from src.data.database import get_session_context
-    from src.data.admin_models import AdminDocument
+    from inference.data.database import get_session_context
+    from inference.data.admin_models import AdminDocument

    with get_session_context() as session:
        statement = select(AdminDocument).where(
@@ -116,8 +116,8 @@ def process_document_autolabel(
    Returns:
        Result dictionary with success status and annotations
    """
-    from src.processing.autolabel_tasks import process_text_pdf, process_scanned_pdf
-    from src.pdf import PDFDocument
+    from training.processing.autolabel_tasks import process_text_pdf, process_scanned_pdf
+    from shared.pdf import PDFDocument

    document_id = str(document.document_id)
    file_path = Path(document.file_path)
@@ -247,7 +247,7 @@ def _save_annotations_to_db(
        Number of annotations saved
    """
    from PIL import Image
-    from src.data.admin_models import FIELD_CLASS_IDS
+    from inference.data.admin_models import FIELD_CLASS_IDS

    # Mapping from CSV field names to internal field names
    CSV_TO_INTERNAL_FIELD: dict[str, str] = {
@@ -480,7 +480,7 @@ def save_manual_annotations_to_document_db(
    pdf_type = "unknown"
    if pdf_path.exists():
        try:
-            from src.pdf import PDFDocument
+            from shared.pdf import PDFDocument
            with PDFDocument(pdf_path) as pdf_doc:
                tokens = list(pdf_doc.extract_text_tokens(0))
                pdf_type = "scanned" if len(tokens) < 10 else "text"
--- a/packages/inference/inference/web/services/inference.py
+++ b/packages/inference/inference/web/services/inference.py
@@ -71,8 +71,8 @@ class InferenceService:
        start_time = time.time()

        try:
-            from src.inference.pipeline import InferencePipeline
-            from src.inference.yolo_detector import YOLODetector
+            from inference.pipeline.pipeline import InferencePipeline
+            from inference.pipeline.yolo_detector import YOLODetector

            # Initialize YOLO detector for visualization
            self._detector = YOLODetector(
@@ -257,7 +257,7 @@ class InferenceService:

    def _save_pdf_visualization(self, pdf_path: Path, doc_id: str) -> Path:
        """Save visualization for PDF (first page)."""
-        from src.pdf.renderer import render_pdf_to_images
+        from shared.pdf.renderer import render_pdf_to_images
        from ultralytics import YOLO
        import io

--- a/packages/inference/inference/web/workers/init.py
+++ b/packages/inference/inference/web/workers/init.py
@@ -4,8 +4,8 @@ Background Task Queues
 Worker queues for asynchronous and batch processing.
 """

-from src.web.workers.async_queue import AsyncTaskQueue, AsyncTask
-from src.web.workers.batch_queue import (
+from inference.web.workers.async_queue import AsyncTaskQueue, AsyncTask
+from inference.web.workers.batch_queue import (
    BatchTaskQueue,
    BatchTask,
    init_batch_queue,
--- a/packages/inference/inference/web/workers/async_queue.py
+++ b/packages/inference/inference/web/workers/async_queue.py
--- a/packages/inference/inference/web/workers/batch_queue.py
+++ b/packages/inference/inference/web/workers/batch_queue.py
--- a/packages/inference/requirements.txt
+++ b/packages/inference/requirements.txt
@@ -0,0 +1,8 @@
+-e ../shared
+fastapi>=0.104.0
+uvicorn[standard]>=0.24.0
+python-multipart>=0.0.6
+sqlmodel>=0.0.22
+ultralytics>=8.1.0
+httpx>=0.25.0
+openai>=1.0.0
--- a/packages/inference/run_server.py
+++ b/packages/inference/run_server.py
@@ -0,0 +1,14 @@
+#!/usr/bin/env python
+"""
+Quick start script for the web server.
+
+Usage:
+    python run_server.py
+    python run_server.py --port 8080
+    python run_server.py --debug --reload
+"""
+
+from inference.cli.serve import main
+
+if __name__ == "__main__":
+    main()
--- a/packages/inference/setup.py
+++ b/packages/inference/setup.py
@@ -0,0 +1,17 @@
+from setuptools import setup, find_packages
+
+setup(
+    name="invoice-inference",
+    version="0.1.0",
+    packages=find_packages(),
+    python_requires=">=3.11",
+    install_requires=[
+        "invoice-shared",
+        "fastapi>=0.104.0",
+        "uvicorn[standard]>=0.24.0",
+        "python-multipart>=0.0.6",
+        "sqlmodel>=0.0.22",
+        "ultralytics>=8.1.0",
+        "httpx>=0.25.0",
+    ],
+)
--- a/packages/shared/requirements.txt
+++ b/packages/shared/requirements.txt
@@ -0,0 +1,9 @@
+PyMuPDF>=1.23.0
+paddleocr>=2.7.0
+Pillow>=10.0.0
+numpy>=1.24.0
+opencv-python>=4.8.0
+psycopg2-binary>=2.9.0
+python-dotenv>=1.0.0
+pyyaml>=6.0
+thefuzz>=0.20.0
--- a/packages/shared/setup.py
+++ b/packages/shared/setup.py
@@ -0,0 +1,19 @@
+from setuptools import setup, find_packages
+
+setup(
+    name="invoice-shared",
+    version="0.1.0",
+    packages=find_packages(),
+    python_requires=">=3.11",
+    install_requires=[
+        "PyMuPDF>=1.23.0",
+        "paddleocr>=2.7.0",
+        "Pillow>=10.0.0",
+        "numpy>=1.24.0",
+        "opencv-python>=4.8.0",
+        "psycopg2-binary>=2.9.0",
+        "python-dotenv>=1.0.0",
+        "pyyaml>=6.0",
+        "thefuzz>=0.20.0",
+    ],
+)
--- a/packages/shared/shared/init.py
+++ b/packages/shared/shared/init.py
--- a/packages/shared/shared/config.py
+++ b/packages/shared/shared/config.py
@@ -7,10 +7,16 @@ import platform
 from pathlib import Path
 from dotenv import load_dotenv

-# Load environment variables from .env file
-# .env is at project root, config.py is in src/
-env_path = Path(__file__).parent.parent / '.env'
-load_dotenv(dotenv_path=env_path)
+# Load environment variables from .env file at project root
+# Walk up from packages/shared/shared/config.py to find project root
+_config_dir = Path(__file__).parent
+for _candidate in [_config_dir.parent.parent.parent, _config_dir.parent.parent, _config_dir.parent]:
+    _env_path = _candidate / '.env'
+    if _env_path.exists():
+        load_dotenv(dotenv_path=_env_path)
+        break
+else:
+    load_dotenv()  # fallback: search cwd and parents

 # Global DPI setting - must match training DPI for optimal model performance
 DEFAULT_DPI = 150
--- a/packages/shared/shared/data/init.py
+++ b/packages/shared/shared/data/init.py
@@ -0,0 +1,3 @@
+from .csv_loader import CSVLoader, InvoiceRow
+
+__all__ = ['CSVLoader', 'InvoiceRow']
--- a/packages/shared/shared/data/csv_loader.py
+++ b/packages/shared/shared/data/csv_loader.py
--- a/packages/shared/shared/data/db.py
+++ b/packages/shared/shared/data/db.py
@@ -9,8 +9,7 @@ from typing import Set, Dict, Any, Optional
 import sys
 from pathlib import Path

-sys.path.insert(0, str(Path(__file__).parent.parent.parent))
-from src.config import get_db_connection_string
+from shared.config import get_db_connection_string


 class DocumentDB:
--- a/packages/shared/shared/exceptions.py
+++ b/packages/shared/shared/exceptions.py
--- a/packages/shared/shared/matcher/init.py
+++ b/packages/shared/shared/matcher/init.py
--- a/packages/shared/shared/matcher/context.py
+++ b/packages/shared/shared/matcher/context.py
--- a/Show More
+++ b/Show More