LoRA-FAQ

TheBFG1324/LoRA-FAQ

3.1

If you are the rightful owner of LoRA-FAQ and would like to certify it and/or have it hosted online, please leave a comment on the right or send an email to dayong@mcphub.com.

LoRA-FAQ Hamilton is a project that fine-tunes GPT-2 using LoRA to provide direct quotes from Alexander Hamilton's Wikipedia page via an MCP server.

LoRA-FAQ Hamilton

LoRA-FAQ Hamilton is a lightweight project that fine-tunes GPT-2 with LoRA on the Alexander Hamilton Wikipedia page.
The model is then exposed as an MCP server tool, so you can query it from any MCP-compatible client and get Hamilton quotes on demand.

Features

  • Fine-tunes GPT-2 with LoRA using Keras/TensorFlow.
  • Uses text extracted from the Alexander Hamilton Wikipedia page.
  • Exposes an MCP tool called answer_faq(question) that returns a related Hamilton quote.
  • Includes a FastAPI service for local testing.

Project Structure

lora_faq_hamilton/
  data/               # Wikipedia source and processed quotes dataset
  lora_train/         # Training scripts
  serve/              # FastAPI inference service
  mcp_server/         # MCP server exposing the answer_faq tool
  LICENSES/           # Wikipedia CC BY-SA 4.0 license
  README.md

Setup

git clone https://github.com/yourname/lora_faq_hamilton.git
cd lora_faq_hamilton
python -m venv .venv && source .venv/bin/activate

# Install dependencies
pip install "tensorflow==2.16.*" "tensorflow-text==2.16.*" \
            "keras>=3.3,<3.6" "keras-hub[nlp]>=0.16" \
            fastapi uvicorn "pydantic<3" httpx mcp

For Apple Silicon (M1/M2/M3):

pip install tensorflow-macos tensorflow-metal "tensorflow-text==2.16.*"

Training

  1. Download and clean the Hamilton Wikipedia page:
python lora_train/build_dataset.py
  1. Fine-tune GPT-2 with LoRA:
python lora_train/train_lora_gpt2.py

Artifacts will be saved to artifacts/lora_gpt2_savedmodel/.

Running the FastAPI Service

uvicorn serve.app:app --reload --port 8008

Test with curl:

curl -X POST localhost:8008/answer \
     -H 'content-type: application/json' \
     -d '{"question":"national bank"}'

Running the MCP Server

python mcp_server/server.py

This exposes a tool:

  • answer_faq(question: str) -> str

Configure your MCP-compatible client (e.g., Claude Desktop, Cursor) to connect to this server.

Example Usage

Input:

Give me a quote about George Washington

Output:

"Hamilton served as George Washington's aide-de-camp..." — (Alexander Hamilton, Wikipedia)
Source: Wikipedia “Alexander Hamilton” (CC BY-SA 4.0).

License & Attribution