giovannioliveira/stem-corpus-sample

STEM Research Corpus — papers with code, telemetry and experiment logs (sample)

A curated sample of recent scientific research artifacts for language-model pre-training. Every record pairs a research paper in full text with the code its authors released; a set of records also includes the measured time-series / telemetry data or experiment logs the research is built on. All content is licensed for commercial use (CC BY 4.0, CC0, MIT, Apache-2.0, BSD and similar).

Part	Records	Tokens	Share
Papers with code	709	309.4M	70.9%
Papers with code and telemetry / time series	10	63.6M	14.6%
Papers with code and experiment logs / traces	5	63.5M	14.5%
Total	724	436.5M	100%

Token counts use the tiktoken o200k_base vocabulary. A pre-training-ready version of the corpus is included in pretrain/ (see below). The composition and quality profile are detailed in REPORT.md.

Contents

items/<arxiv_id>/                         papers with code
    paper.txt       full text of the paper
    code.txt        the repository as text: file tree, then every source file and its documentation/configuration
    repo.tar.gz     snapshot of the repository at the recorded commit
    meta.json       paper and repository metadata, licenses, quality signals, token counts
datasets/<kind>/<field>/<record>/         papers with code and data (kind: telemetry_timeseries | experiment_logs)
    paper.txt       full text of the paper describing the data
    code.txt        code that produced or analyses the data
    data.txt        the dataset's text content (time series, logs, traces) ready for training
    raw.tar | data/ the dataset's original files as published
    repo.tar.gz     repository snapshot, when the code is a separate repository
    meta.json       dataset, paper and code metadata, licenses, token counts
pretrain/                                 pre-training shards (gzip JSONL) and their manifest
manifest_paper_code.jsonl                 one line per paper-with-code record
manifest_datasets.jsonl                   one line per data record
REPORT.md                                 composition and quality profile

Text formats

• paper.txt: text from the publisher/arXiv PDF text layer, one paper per file.
• code.txt and data.txt: each file is introduced by ===== <path> =====; notebooks are rendered as code and markdown cells; compressed files are decompressed; Parquet/Arrow rows appear as JSON lines. Data records carry an even sample of their files (up to 10M tokens each) in data.txt, with the complete originals alongside.

Pre-training shards (pretrain/)

24 gzip JSONL shards, one document per line: a text field plus metadata (part, record, licenses, languages, gpu_workflow, tokens, …). Each document joins a full paper with its complete repository, in the repository-level layout used for code-model pre-training:

<paper title="…" id="arXiv:…">
…full paper…
</paper>
<repo_name>owner/repo
<file_sep>README.md
…
<file_sep>src/module.py          (sources in dependency order, tests last)
…
<data source="…">              (data records)
<file_sep>path
…
</data>

Metadata

Each record's meta.json and manifest line give identifiers and links (arXiv id, DOI, Zenodo / Hugging Face / GitHub), title, authors, abstract or description, arXiv category or scientific field, publication month, venue, licenses of the paper, code and data, repository URL and commit, quality signals and token counts per file.

Licensing and attribution

Papers: CC BY 4.0 or CC0. Code: the repository's own license (MIT, Apache-2.0, BSD, ISC, CC0 and similar). Data: the dataset's published license (CC BY 4.0, CC0, MIT, Apache-2.0). Each record states its licenses in meta.json; CC BY works are attributed by citing the paper and the dataset or repository listed there.