gzb666/llvm_training_data

LLVM Commit Training Dataset

Description

This dataset contains 256,055 code-modification training samples extracted from the LLVM project git history. Each sample pairs a natural-language instruction (derived from the commit message) with the before/after code state, suitable for supervised fine-tuning of LLMs on C++ code refactoring, bug fixing, and feature implementation tasks.

Dataset Structure

Each line in llvm_training_data.jsonl.gz is a JSON object with the following fields:

Field	Type	Description
instruction	string	Natural-language task description derived from commit subject + body. Includes file paths and feature/refactor description.
input	string	Code state before the commit (extracted from parent commit). Empty for new-file creation tasks.
output	string	Code state after the commit. May contain /* --- N lines omitted --- */ markers where unchanged context was folded.
file_paths	list[string]	Source file paths touched by this record (1-2 files per record).
commit_hash	string	Full 40-char SHA-1 of the source LLVM commit.
author	string	Commit author name.
author_date	string	Commit date in YYYY-MM-DD format.
task_type	string	Always code_modification for this file.

Output Markers

The output field uses two marker types to keep records within a reasonable context window (~250 lines max):

• /* --- N lines before/after modification point omitted --- */ — folded unchanged code. The original code is balanced; the marker replaces lines that contain braces, so naive brace counting on output alone may show imbalance. This is expected and not a defect.
• /* --- separate code section --- */ — separates multiple code blocks (e.g., a .h declaration and its .cpp implementation) combined into one training record.

Statistics

• Total records: 256,055
• Source commits: 204,622 unique LLVM commits
• Unique authors: 5,255
• Date range: 2001-06-07 to 2026-06-08
• Output length: median 49 lines, mean 85.4 lines, max 1021 lines
• New-file records (empty input): 6,022 (2.4%) — "implement new feature" tasks
• Multi-file records: 26,300 (10.3%) — paired header/implementation changes

File extension distribution

Extension	Count
.cpp	205,843
.h	70,286
.cc	4,763
.inc	1,272
.hpp	190

Component distribution (top-level LLVM subprojects)

Component	Count
llvm	135,754
clang	65,256
lldb	19,148
mlir	16,917
lld	10,820
compiler-rt	8,597
flang	8,181
clang-tools-extra	6,194
bolt	2,310
polly	2,241

Quality Filtering

Each record passed the following quality gates:

1. Pipeline extraction — AST-based (tree-sitter) function block extraction with regex fallback, three-layer before/after pairing (exact name → fuzzy → positional fallback), folding of unchanged context, and multi-window truncation.
2. Quality check — passed 17+ heuristic checks including brace balance, identifier retention, function pairing consistency, size mismatch, severe duplication, and context anchor preservation.
3. Post-filter — a final brace-integrity check that rejects records with genuinely broken braces (not explainable by folding markers). 15,918 records (5.85% of pre-filter PASS) were downgraded to a separate ARBITRATE set (not included in this upload).

After post-filtering, a 200-sample audit showed 0% broken-brace records.

Source

• Repository: llvm/llvm-project
• Extraction pipeline: Custom Python pipeline processing 583,490 non-merge commits from the LLVM git history.
• Toxic commit elimination: 9,754 reverted commits and their targets were excluded.
• Deduplication: Three-stage dedup (exact match → homogenization Jaccard → MinHash LSH).

Intended Uses

• Fine-tuning LLMs for C++ code modification tasks
• Studying commit-level code transformations
• Benchmarking code generation on real-world refactoring scenarios

Limitations

• Folding markers: The omitted markers mean output is not always directly compilable. Models trained on this data should learn to interpret the markers as "unchanged context elided," not as literal code.
• LLVM-specific patterns: Code heavily features LLVM APIs (StringRef, SmallVector, DenseMap, etc.) and conventions (2-space indent, llvm:: namespace). May not generalize to non-LLVM C++ without domain adaptation.
• Historical bias: Older commits (2001-2010) reflect coding standards and APIs that may have since been deprecated or refactored in LLVM.
• Extracted subset: Each record is a windowed view of the commit, not the complete diff. Large multi-file commits are represented by 1-2 representative files.

License

Apache License 2.0 with LLVM Exception — same as the LLVM project source.

The dataset is derived from LLVM commit history. Each record's commit_hash field allows tracing back to the original upstream commit for attribution.

Citation

If you use this dataset, please cite the LLVM project:

@misc{llvm-project,
  title        = {The LLVM Compiler Infrastructure},
  author       = {{LLVM Project Contributors}},
  howpublished = {\url{https://github.com/llvm/llvm-project}},
  year         = {2003--2026},
}