Upload folder using huggingface_hub

a4be7e8 verified 9 days ago

2.67 kB

license: apache-2.0
base_model: mlx-community/Qwen2.5-Coder-7B-Instruct-4bit
tags:
  - mlx
  - lora
  - qwen
  - commit-message
  - conventional-commits
  - git
  - chinese

git-ai-commit-sft

A LoRA adapter fine-tuned on Qwen2.5-Coder-7B-Instruct-4bit to generate Chinese Conventional Commits commit messages from git diffs.

Built for git-ai-commit, an IntelliJ plugin that uses LLMs to generate commit messages.

Model

Base model: mlx-community/Qwen2.5-Coder-7B-Instruct-4bit
Fine-tuning: LoRA (rank 8, 16 layers, 11.5M trainable params)
Format: MLX adapter (safetensors)

Training Data

548 high-quality commit messages from two real-world repositories (one Java backend, one Go microservice), filtered to Chinese-only Conventional Commits format. Each training sample pairs a git diff (processed through the plugin's exact runtime pipeline — GitDiffFilter + PromptBuilder) with the corresponding human-written commit message.

Evaluation

Metric	Before	After
Conventional Commits rate	96%	100%
Chinese rate	100%	100%
Single-line rate	100%	100%
Mean similarity to reference	0.339	0.546
Mean output length	51 chars	27 chars

Usage

# Install mlx-lm
pip install mlx-lm

# Download and load
python -m mlx_lm.generate \
  --model mlx-community/Qwen2.5-Coder-7B-Instruct-4bit \
  --adapter-path yisuiban/git-ai-commit-sft \
  --prompt "你是一位资深工程师，擅长根据 git diff 生成一句中文提交信息。..."

From Python:

from mlx_lm import load, generate
from mlx_lm.sample_utils import make_sampler

model, tokenizer = load(
    "mlx-community/Qwen2.5-Coder-7B-Instruct-4bit",
    adapter_path="yisuiban/git-ai-commit-sft"
)

prompt = tokenizer.apply_chat_template(
    [{"role": "user", "content": "你的提示词..."}],
    tokenize=False, add_generation_prompt=True
)
response = generate(model, tokenizer, prompt=prompt, max_tokens=64,
                    sampler=make_sampler(temp=0.0))

For Ollama deployment, see the training repository for the full Modelfile and export pipeline.

Training Details

Hardware: Apple M1 Max (64GB unified memory)
Framework: MLX LoRA (mlx-lm 0.31.3)
Optimizer: Adam, learning rate 1e-5
Batch: 1 × gradient accumulation 8 (effective batch 8)
Steps: 250 (~4 epochs over 473 training samples)
Max sequence length: 4096 tokens
Training time: ~20 minutes on M1 Max