Qwen2.5-7B Math Fine-tuning System

This repository provides a structured training and evaluation pipeline for fine-tuning Qwen/Qwen2.5-7B-Instruct on grade-school math tasks with LoRA.

Quick Start

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
python scripts/generate_data.py --num_samples 200 --output data/examples/train.json --validate
python scripts/train.py --data_path data/examples/train.json --quick_test
python scripts/evaluate.py --model_path outputs/models/default/final --benchmark gsm8k

Docker Images

This project builds 3 Docker images on each GitHub release:

Image	Purpose
ghcr.io/elloloop/primary-math-finetuning/train	Fine-tune Qwen2.5-7B with LoRA on RunPod GPU Pods
ghcr.io/elloloop/primary-math-finetuning/inference	Serve the model via FastAPI + vLLM (RunPod Serverless)
ghcr.io/elloloop/primary-math-finetuning/eval	Benchmark against GSM8K on RunPod GPU Pods

See docs/DEPLOYMENT.md for the full RunPod walkthrough.

Docs

• docs/DEPLOYMENT.md — RunPod deployment guide (start here)
• docs/TRAINING.md
• docs/EVALUATION.md
• docs/DATA_PREPARATION.md
• docs/API.md
• docs/TROUBLESHOOTING.md