Ryze Fine-Tune Pipeline

A comprehensive pipeline for fine-tuning LLMs on domain-specific documents. Converts PDF files into training data via OCR, generates QA pairs, and runs two-stage LoRA training (SFT → GRPO).

PDF Library  →  OCR + QA Generation  →  RL Training  →  Fine-tuned Model

• Upload PDF files
• OCR extraction (GLM-OCR, Tesseract, PaddleOCR, EasyOCR, DocTR)
• Automatic QA pair generation via LLM (Qwen3, GPT-4o, Claude, Llama, DeepSeek)

• Supervised Fine-Tuning with LoRA on generated QA dataset
• Configurable: base model, learning rate, epochs, batch size

• Group Relative Policy Optimization on SFT-merged model
• Reward functions: BLEU, ROUGE-L, BERTScore, custom
• LoRA adapter auto-merges to produce final model

Ryze/
├── src/
│   ├── data/       # PDF OCR, QA dataset generation
│   ├── rl/         # SFT and GRPO trainers, LoRA utilities
│   ├── eval/       # Evaluation metrics and benchmarks
│   ├── ui/         # Gradio interface
│   └── figure/     # Figure extraction utilities
├── configs/        # Training configuration files
├── scripts/        # Utility scripts
├── raze_traing_demo/  # Static pipeline demo (index.html)
└── demo/           # Streamlit demos

# Install Tesseract OCR
sudo apt-get install tesseract-ocr tesseract-ocr-eng

# Install Python dependencies
pip install -r requirements.txt

./start.sh

Or manually:

python launch_app.py

Then open http://localhost:7860 in your browser.

Edit configs/default_config_v2.json to customize OCR settings, training hyperparameters, model paths, and output directories.

1. Data Processing tab — Upload PDFs, select OCR engine and QA LLM, run processing
2. Model Training tab — Configure SFT LoRA, then GRPO LoRA; models auto-merge between stages
3. Evaluation tab — Benchmark final model

Apache License 2.0