abhifdsdf/cuad-raft

⚖️ CUAD RAFT — Legal Contract Q&A

Fine-tuned Qwen2.5-3B via RAFT (Retrieval-Augmented Fine-Tuning) on CUAD legal contracts.

How to Deploy

1. Create a new HF Space

Go to https://huggingface.co/new-space and fill:

• —Space Name: cuad-raft-demo (or any name)
• —License: Apache-2.0
• —Space SDK: Gradio
• —Hardware: Free CPU (2 vCPU · 16GB · 30m sleep) — sufficient for the Q4KM GGUF (~2GB)

2. Upload files

Push these 3 files to the Space:

# Clone the empty space (use your own space name)
git clone https://huggingface.co/spaces/YOUR_USERNAME/cuad-raft-demo
cd cuad-raft-demo

# Copy files from this repo
cp path/to/app.py .
cp path/to/requirements.txt .
cp path/to/README.md .

git add .
git commit -m "Initial deploy: CUAD RAFT GGUF demo"
git push

3. Set model cache (optional)

Add a Space Secret (Settings → Repository Secrets):

MODEL_CACHE_DIR = /data/.cache

This keeps the 2GB GGUF cached across restarts instead of re-downloading.

4. Wait for build (~5 min)

The Space builds on first push. llama-cpp-python compiles C++ code, so this takes longer on the first push. Subsequent pushes are instant.

API Usage

Once deployed, the API is at:

POST https://YOUR-SPACE.hf.space/api/generate
Content-Type: application/json

{
  "question": "What is the governing law?",
  "context": "This agreement shall be governed by the laws of Delaware.",
  "temperature": 0.0,
  "max_tokens": 256
}

Response:

{
  "answer": "Delaware law.",
  "latency_seconds": 2.34
}

Files

Model

• —GGUF: abhifdsdf/qwen2.5-3b-cuad-raft-gguf
• —Format: Q4KM (~2GB, fits 16GB RAM)
• —Base: Qwen2.5-3B-Instruct
• —Fine-tuned via: RAFT on CUAD dataset