--- license: other license_name: mixed license_link: LICENSE.md language: [en] pretty_name: length-extrapolating-architecture training data --- # Training data for the RSI Bench task length-extrapolating-architecture 1,010,115,101 tokens (Mistral-7B tokenizer, revision 27d67f1b5f57), stored as `uint16` in `train.bin`. Built by `data_build_v2.py` on 2026-10-13; `manifest.json` records every source revision, filter and hash. ## Files | File | What it is | |---|---| | `train.bin` | the token stream; long documents start at multiples of 2,048 tokens, end-of-text id 2 between documents | | `segments.npy` | one row per laid-out segment: start, length, source, document length, kind, row of `docs.jsonl` | | `docs.jsonl` | one row per document: source, reference (repository, arXiv ID, PG-19 ID, FineWeb-Edu ID), licence, tokens, sha256 of its tokens | | `repo_pins.json` | each repository: commit (the last before 2025-01-01), commit date, licence, file count | | `notices.tar.gz` | each repository's LICENSE, COPYING, NOTICE and COPYRIGHT files, as found at the pinned commit | | `manifest.json` | sources, revisions, mix, filters, statistics, hashes | | `dev/` | the task's development set (repositories and arXiv papers from May to September 2026): token files, character offsets, texts, hard-token masks, `index.json` (licence, commit or arXiv ID, scoring length), `MANIFEST.json` (sha256 of every file) and `notices/` (each repository's licence files) | ## Sources and licences Each part keeps its own licence. This dataset as a whole is offered under those licences and no other. | Source | Share | Licence | |---|---|---| | FineWeb-Edu (`HuggingFaceFW/fineweb-edu` at `87f09149ef47`) | 35% | ODC-By 1.0; credit to Hugging Face (FineWeb-Edu) and Common Crawl | | GitHub repositories (`repo_pins.json`) | 30% | per repository: MIT, Apache-2.0, BSD-2-Clause, BSD-3-Clause or ISC; licence and notice texts in `notices.tar.gz` | | arXiv papers (`common-pile/arxiv_papers` at `963fe980c55b`) | 15% | per paper, CC BY or CC0 (listed in `docs.jsonl`); credit to each paper's authors | | PG-19 books (`deepmind/pg19` at `4d28bd77e669`) | 20% | Apache-2.0 dataset; the books are in the US public domain | Documents by source and licence: | Source | Licence | Documents | |---|---|---| | book | Apache-2.0 dataset, public-domain texts | 3,823 | | code | apache-2.0 | 2,020 | | code | bsd-2-clause | 87 | | code | bsd-3-clause | 287 | | code | isc | 23 | | code | mit | 4,193 | | paper | Creative Commons - Attribution - https://creativecommons.org/licenses/by/3.0/ | 421 | | paper | Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/ | 6,218 | | paper | Creative Commons Zero - Public Domain - https://creativecommons.org/publicdomain/zero/1.0/ | 266 | | web | ODC-By 1.0 | 311,531 | ## Changes made to the sources - Text was tokenized; line endings became "\n"; the literal text ``, `` and `` was encoded as plain text. - Repositories: code files only (rules in the build's README), source folders first, joined with a `# ===== file: =====` line before each file. Licence and notice files were kept separately (`notices.tar.gz`). - Every document was cut at its first 65,536 tokens. - FineWeb-Edu pages from Project Gutenberg were removed, and documents that copy the task's evaluation documents were removed. ## Takedown To ask for a document's removal, open a discussion on this dataset's page with the reference from `docs.jsonl`.