CommitPackFT: 2GB of Instruction-Filtered Code Commit Messages

CommitPackFT is a 2GB collection of high-quality code commit messages filtered to resemble natural language instructions, containing between 100,000 and 1,000,000 records. Developed by BigCode and released in August 2023, it serves as a fine-tuning variant of the larger CommitPack dataset for instruction-following tasks. The data is linked to the research findings in Arxiv paper 2308.07124.

Use Cases

• Fine-tuning Large Language Models (LLMs) for code generation using commit messages as natural language prompts
• Instruction tuning for models to interpret developer intent from commit-style descriptions
• Training models to automate the generation of commit messages from code diffs

Strengths

• 2GB of high-quality filtered text-code pairs
• MIT licensed for commercial and research use
• Specifically curated for instruction-following patterns rather than raw commit logs
• Validated through the OctoPack research methodology (Arxiv 2308.07124)

Limitations

• Specific column schema is not explicitly documented in the source metadata
• Potential geographic bias toward US-based regions as indicated by metadata tags
• Limited to commit-style interactions which may not cover all software engineering instruction types

Provenance

Source

BigCode (Arxiv 2308.07124)

Collection Method

Algorithmic filtering of the CommitPack dataset to isolate messages resembling natural language instructions

Freshness

Last updated August 2023; reflects code repository states up to the point of the original CommitPack collection.

Geography

Global