williamTLmiller/python-documentation-rag-corpus

Python Library Documentation Corpus for RAG & Continued Pretraining (Version-Aware)

Release name: MatrixNorm DomainPacks software_docs v1 (free sample)

Version-aware software documentation for continued pretraining and retrieval: the official docs of Python 3.13 and 3.12, Flask, Click, Jinja, Werkzeug and attrs, read from sha256-pinned release archives, normalized, split into heading-scoped chunks that keep code blocks whole, and de-duplicated across versions (3.12 keeps only the text that differs from 3.13). Each pretraining row carries its version context and the exact archive and character range it came from. Retrieval rows ask about a specific API symbol or section in a specific version, with the exact supporting chunk as the positive, same-document hard negatives that never mention the target, and unanswerable queries for plausible API names that exist nowhere in that project's docs.

This is the free public sample: whole train groups only (about 10% of them). Validation and test rows stay in the private full release so they remain unseen. Splits below are therefore train only.

Rows

View	train
domain_pretraining	453
domain_retrieval	452

from datasets import load_dataset
ds = load_dataset("williamTLmiller/matrixnorm-domainpacks-software-docs-v1-sample", "domain_pretraining")

Each view is also stored as JSON Lines next to the Parquet file, with identical rows. Nested fields (messages, citations, source_refs, tables) are native lists and structs in both.

Views. domain_pretraining: text (an exact slice of the normalized source file), language, profile, version_context (for example Python 3.13.0) and source_refs (archive URL, version plus archive sha256, license, and path#codepoints:start:end). domain_retrieval: query, positives, hard_negatives and answerability (answerable or unanswerable). For RAG evaluation, an unanswerable query has no positive, and the hard negative is the closest real passage, which does not answer it.

How it was built

Sources are pinned by exact release-archive URL and sha256 (python.org and PyPI release files). Only the documentation directory of each archive is read, as text; nothing is executed. Python code examples are syntax-checked with ast.parse and recorded as pass or fail; they are never run. Chunks target 4,000 characters (maximum 8,000) with 300-character overlap when a long section must be split, and splits only fall on blank lines outside literal blocks, code blocks and tables. A section that cannot be split under the maximum is quarantined, never truncated. Exact duplicates (after whitespace normalization) keep the newest version. Groups = the same doc path across versions, merged with any documents that share a duplicate chunk, so the same text never lands on both sides of a split. Token counts use a pinned GPT-2-style pre-tokenizer regex (matrixnorm-pretok-regex-v1), not a learned vocabulary.

Validation

Every record is checked before it is written: schema, cross-references, split membership, and the dataset-specific checks described above. A family that fails any check is quarantined as a whole, never silently dropped or partially kept.

Build stats: {"chunks_exclude": 2512, "chunks_include": 4408, "chunks_quarantine": 28, "documents": 1156, "examples_not_run": 1049, "examples_syntax_fail": 1506, "examples_syntax_pass": 8380, "symbols": 17525}

Trainability baseline

Measured on the full release's held-out splits (this sample is train only). TF-IDF retrieval (vectorizer fit on train pretraining text only), ranking each held-out query against that split's pooled positives and hard negatives:

Split	Answerable queries	Pool	Recall@1	Recall@5	MRR	Random Recall@5
validation	44	30	0.250	0.523	0.374	0.167
test	31	29	0.484	0.613	0.563	0.172

The held-out sets are small (see the pilot note above), so treat these as a sanity check that the pairs are learnable, not as a benchmark.

Limitations

• Seven documentation trees only; Python dominates the token count. Retrieval positives come from the symbol's own directive or section, so paraphrased or multi-hop questions are not covered.
• reStructuredText and MyST markup is kept as written (it is the source text); strip it if you need plain prose.
• Python 3.12 rows are only the chunks whose text differs from 3.13.

Licensing

CC BY-NC 4.0 (see LICENSE and NOTICE). The license covers only the MatrixNorm packaging layer; the documentation text keeps its own PSF-2.0 / BSD-3-Clause / MIT license, copied in source-license-copies/. Commercial licensing: energea@energeatech.com.

Citation

MatrixNorm DomainPacks software_docs v1. williamTLmiller, MatrixNorm, 2026. https://huggingface.co/datasets/williamTLmiller/matrixnorm-domainpacks-software-docs-v1-sample