| --- |
| license: apache-2.0 |
| language: |
| - en |
| pretty_name: Mathlib Refactor History |
| size_categories: |
| - 10K<n<100K |
| tags: |
| - lean |
| - mathlib |
| - theorem-proving |
| - software-history |
| - refactoring |
| configs: |
| - config_name: refactor_pair_catalog |
| default: true |
| data_files: |
| - split: full |
| path: data/refactor_pair_catalog.jsonl |
| - config_name: commits |
| data_files: |
| - split: full |
| path: data/commits/*.jsonl.gz |
| - config_name: statement_scans |
| data_files: |
| - split: full |
| path: data/statement_scans/*.jsonl.gz |
| --- |
| |
| # Mathlib Refactor History |
|
|
| This release is a source-level history of public declarations in |
| [`leanprover-community/mathlib4`](https://github.com/leanprover-community/mathlib4), |
| pinned to commit `e72c1e277f31441626621f7d0c7207862fc25569`. It connects a complete |
| commit index to statement-change scans and lossless before/after bundles for |
| refactor-shaped historical events. |
|
|
| The package contains 157.7 MiB of compressed/derived data in |
| 17 data files. Exact byte counts and both compressed and |
| canonical SHA-256 digests are recorded in `RELEASE_MANIFEST.json`. |
|
|
| ## Data layers |
|
|
| | Layer | Records | Hub access | Meaning | |
| | --- | ---: | --- | --- | |
| | `refactor_pair_catalog` | 13,905 | Default viewer configuration | Compact lookup table with provenance, screening status, declaration counts, shard location, and per-record hash | |
| | `commits` | 33,999 | Viewer configuration | Every commit reachable from the frozen snapshot, including first-parent metadata, changed Lean files, PR reference, and historical Lean toolchain when present | |
| | `statement_scans` | 33,999 | Viewer configuration | Source-structure changes for the same commits, including public declaration events and high-recall structural signals | |
| | Full refactor pairs | 13,905 | `data/refactor_pairs/*.jsonl.gz` | Complete nested before/after declaration bundles with source statements, context, alignment, provenance, and screening metadata | |
|
|
| The full pair payload has 11 record-aligned shards. It is kept |
| outside the hosted viewer configurations because 46 valid records are larger than |
| 1 MiB and the largest is about 50 MiB. Nothing is removed: use the catalog's |
| `full_record_file` and `record_index_in_shard` fields to locate a full record. |
|
|
| ## Important quality status |
|
|
| This is a **high-recall structural release, not an expert-verified gold dataset**. |
|
|
| - 13,900 pair boundaries are `provisional_structural`; 5 use curated pilot |
| overrides. |
| - AI screening labels are triage metadata. They do not establish mathematical |
| equivalence, shared intent, quality, or preference. |
| - Proof-body-only changes to theorems and lemmas are ignored by the statement |
| scanner. Definition/abbreviation value changes are retained. |
| - 76 commit-level scan records contain truncated event previews; their summaries |
| remain complete, and full pair materialization recomputed structural events from |
| the parsed-blob cache. |
| - There are no train/validation/test splits. The split is named `full` to avoid |
| implying that the records are ready-made training examples. |
|
|
| Do not report model performance on these pairs as gold refactoring accuracy |
| without an independent human review and a documented filtering protocol. |
|
|
| ## Load the viewer-friendly tables |
|
|
| ```python |
| from datasets import load_dataset |
| |
| repo_id = "YOUR_NAMESPACE/mathlib-refactor-history" |
| |
| catalog = load_dataset( |
| repo_id, "refactor_pair_catalog", split="full" |
| ) |
| commits = load_dataset(repo_id, "commits", split="full") |
| scans = load_dataset(repo_id, "statement_scans", split="full") |
| ``` |
|
|
| ## Stream the lossless full pair records |
|
|
| Download the snapshot and let the JSON loader decompress the Gzip shards: |
|
|
| ```python |
| from pathlib import Path |
| |
| from datasets import load_dataset |
| from huggingface_hub import snapshot_download |
| |
| repo_id = "YOUR_NAMESPACE/mathlib-refactor-history" |
| snapshot = Path(snapshot_download(repo_id, repo_type="dataset")) |
| pair_files = [str(path) for path in sorted( |
| snapshot.glob("data/refactor_pairs/*.jsonl.gz") |
| )] |
| |
| pairs = load_dataset( |
| "json", |
| data_files={"full": pair_files}, |
| split="full", |
| streaming=True, |
| ) |
| first_pair = next(iter(pairs)) |
| ``` |
|
|
| For a dependency-light workflow, use `gzip -dc` and parse one JSON object per |
| line. Run `python3 scripts/verify_release.py` after downloading to check every |
| file, decompressed table digest, record count, and pair-to-catalog mapping. |
|
|
| ## Schemas and identifiers |
|
|
| - `schemas/commit_history.schema.json` describes commit records. Primary key: |
| `commit`. |
| - `schemas/statement_history_scan.schema.json` describes scan records. Primary |
| key: `commit`. |
| - `schemas/pair.schema.json` describes full pair records. Primary key: `pair_id`. |
| - `schemas/refactor_pair_catalog.schema.json` describes the compact catalog. |
|
|
| All commit- and scan-level tables have one record per reachable commit and join on |
| `commit`. Pair provenance contains `before_commit` and `after_commit`; the release |
| does not assert that all provisional bundles express one minimal mathematical |
| intent. |
|
|
| ## Construction and reproducibility |
|
|
| The history was indexed and scanned from the frozen Mathlib snapshot. The source |
| parser is `lean-source-declarations-v0.4`; the statement scanner is |
| `source-structure-v0.1`. The package retains the canonical uncompressed SHA-256 of |
| each source JSONL in `RELEASE_MANIFEST.json`, so repackaging does not change data |
| identity. |
|
|
| The package deliberately excludes local Git/build caches, SQLite parsed-blob |
| caches, archived parser versions, duplicated benchmark materializations, |
| discussion bodies, consumer-graph build trees, and pending natural-language or |
| visibility drafts. See `docs/decisions/0001-hugging-face-release-package.md` for |
| the scope rationale. |
|
|
| ## Personal and sensitive information |
|
|
| Commit records retain public Git author names and commit subjects for provenance; |
| they do not contain author email addresses. Pair records contain excerpts of the |
| publicly licensed Mathlib source and public GitHub URLs. An automated pre-release |
| scan found no local absolute workspace paths, Hugging Face/GitHub access tokens, |
| PEM private keys, or AWS access-key patterns in the selected canonical sources. |
| Automated scanning cannot prove the absence of all sensitive content. |
|
|
| GitHub and Zulip discussion bodies are intentionally not included in this |
| release. Please use public identities responsibly and follow applicable deletion |
| or correction requests. |
|
|
| ## License and attribution |
|
|
| The release is distributed under Apache License 2.0. It includes mechanically |
| derived metadata and source excerpts from Mathlib, which is also distributed |
| under Apache License 2.0. See `LICENSE` and `NOTICE.md`. Repository links and |
| commit hashes are retained for provenance and attribution. |
|
|
| ## Citation |
|
|
| Until an archival DOI is assigned, cite the Hugging Face dataset revision you |
| used together with the pinned Mathlib snapshot. Also cite Mathlib according to |
| the guidance in the upstream Mathlib repository. |
|
|
