SFT Collector Project

A Python web application for dataset collection using Flask, SQLite, and Gunicorn. This project was created largely thanks to ChatGPT 4. See https://chat.openai.com/share/1d54fad4-a65d-4442-85b6-0e4e98cd9f4f for the conversation that led to this project.

Prerequisites

Python 3.7+
pip (Python package manager)

Installation

Clone or download the project files.
Create a virtual environment to isolate your project dependencies:

python -m venv venv
Activate the virtual environment:
- On Windows: venv\Scripts\activate
- On macOS/Linux: source venv/bin/activate
Install the required packages:

pip install -r requirements.txt
Download a small language model of your choice for "Ask a language model":

litgpt download --repo_id "microsoft/phi-1_5"

Usage

Run the Application

Start the application server using Flask local server:

python sft_collector:app

sft_collector is the name of your Python script (without the .py extension).
app is the name of the Flask app instance inside sft_collector.py.
Somehow I cannot get lit-gpt work with gunicorn.
Put PYTORCH_ENABLE_MPS_FALLBACK=1 in front, if you are running it on Mac with Apple silicon.

Using the collected dataset

Convert it to .json file

This repository comes with a script that returns a json file.

python produce_json.py --dataset_name YOUR_DATASET_NAME --sqlie_file datasets.db --json_file YOUR_DATASET_NAME.json

Use it with LitGPT

Finetune your favourite model with the collected dataset:

litgpt finetune lora --checkpoint_dir ./checkpoints/microsoft/phi-1_5 --data JSON --data.json_path ./YOUR_DATASET_NAME.json --data.val_split_fraction 0.1 --out_dir out/my_model/

kyunghyuncho/sft_collector