Dataset Card for Code Corpus code-v1

Dataset Description

Dataset Summary

Code Corpus code-v1 is a token-balanced collection of source code, commit messages and diffs, programming Q&A, and technical text. It is intended for language-model pretraining and continued pretraining. All examples use a common JSONL schema with per-record provenance, license metadata, exact token counts, and content hashes.

The released dataset contains 2 billion tokens, 2,541,336 documents, and 46 gzip-compressed shards totaling 2.91 GB (2,905,546,899 bytes; 2.71 GiB). Exact source totals and the complete shard inventory are recorded in manifest.json.

Supported Tasks

The dataset is designed for causal language modeling over code and technical text. It is not an instruction-tuning dataset and does not include labels, preference pairs, or a held-out evaluation split.

Languages

The corpus contains many programming languages and predominantly English natural-language content. Other human languages may occur in source comments, commit messages, repository content, and upstream technical-text collections. The language field describes a programming language when upstream metadata is available; it is not a human-language tag.

Dataset Structure

Data Instances

Each line in a decompressed shard is an independent JSON object:

{
  "id": "29266961adb8c59a77001330377561454f9d82e5dadfeff7445a14ae147ac453",
  "text": "// source code, technical text, or a commit message and diff",
  "source": "stackv2_edu",
  "source_id": "6396980ea790ba1a35f35ff6269bb0d8d0b58ddf",
  "language": "JavaScript",
  "license": "MIT",
  "url": "https://raw.githubusercontent.com/...",
  "repository": "owner/repository",
  "path": "/path/to/file.js",
  "created": "2020-03-02 08:38:22",
  "metadata": {},
  "content_sha256": "72f9484f64be497baa6bd276fbfdbed13e43735445f0aedb635907314d223ceb",
  "chunk_index": 0,
  "token_count": 599
}

CommitPack examples format text as a natural-language commit message followed by a unified diff.

Data Fields

Field	Type	Description
id	string	Stable SHA-256 record identifier.
text	string	Model training text.
source	string	Source name matching a recipe entry and shard directory.
source_id	string or null	Upstream example identifier.
language	string or null	Upstream or detected programming language.
license	string or null	Upstream license label for the example.
url	string or null	Original or raw-content URL when available.
repository	string or null	Repository name when available.
path	string or null	Path within the repository when available.
created	string or null	Upstream timestamp; formatting varies by source.
metadata	object	Source-specific provenance and attributes.
content_sha256	string	SHA-256 hash used for exact content deduplication.
chunk_index	integer	Zero-based index when a long document is split.
token_count	integer	Exact text token count using the configured tokenizer.

Missing provenance is represented as null, not inferred. Consumers should allow source-specific keys in metadata.

Data Splits

There is one train split. The recipe targets the following token mixture:

Source	Upstream dataset	Share	Nominal tokens
stackv2_edu	common-pile/stackv2_edu_filtered	65.0%	1,300,000,000
stackv2_production	common-pile/stackv2	10.0%	200,000,000
commitpack	bigcode/commitpackft	8.0%	160,000,000
stackexchange_programming	common-pile/stackexchange_filtered	7.0%	140,000,000
arxiv_cs_math	common-pile/arxiv_papers_filtered	5.0%	100,000,000
repository_docs	common-pile/stackv2_edu_filtered	2.9%	58,000,000
python_peps	common-pile/python_enhancement_proposals_filtered	0.1%	2,000,000
wikimedia	common-pile/wikimedia_filtered	2.0%	40,000,000

If an upstream source is exhausted before filling its nominal allocation, the shortfall is assigned to stackv2_edu. The manifest records the realized source proportions; recipe.json records the target mixture and selection rules.

Loading the Dataset

After upload to the Hub:

from datasets import load_dataset

dataset = load_dataset("owenqwenllmwine/code-v1", split="train")

To load this directory locally:

dataset = load_dataset(
    "json",
    data_files="data/*/train-*.jsonl.gz",
    split="train",
)

The card's default configuration maps all compressed JSONL shards to the train split.

Dataset Creation

Curation Rationale

The recipe emphasizes educational source code while retaining smaller amounts of production code, code changes, programming discussions, repository documentation, papers, standards, and general text. Source shares are allocated by tokenizer tokens rather than document count or compressed size.

Source Data

The data is derived from the upstream Hugging Face datasets linked above. The recipe uses seed 42, a 10,000-example streaming shuffle buffer, a minimum document length of 32 tokens, and a maximum chunk length of 32,768 tokens. Token counts use Qwen/Qwen3.5-0.8B-Base.

Processing includes:

• normalization into the common record schema;
• exact cross-source deduplication;
• long-document chunking;
• redaction of private-key blocks and common live-token patterns; and
• filtering of empty, repetitive, minified, generated, and common vendor-path content.

Educational code excludes documentation and notebooks. Production code additionally excludes common text, lock, source-map, and tabular files. Programming Q&A is restricted to selected programming-focused Stack Exchange sites. The exact filters are recorded in the recipe.

Annotations

No new labels or human annotations were created. Language, license, provenance, and other attributes are retained or normalized from upstream metadata.

Personal and Sensitive Information

Public source code and technical discussions can contain names, email addresses, usernames, URLs, credentials, or other personal and sensitive information. Automated secret-pattern redaction reduces some credential exposure but is not comprehensive. No dedicated PII removal or privacy audit has been performed.

Considerations for Using the Data

Intended Use

Appropriate uses include language-model pretraining, continued pretraining, and research on code and technical text. Users should retain provenance where needed for attribution and perform additional filtering, decontamination, or safety review appropriate to their use case.

Out-of-Scope Use

The corpus should not be treated as a source of verified, secure, up-to-date, or executable code. It is not suitable by itself for model evaluation, supervised instruction tuning, or applications that require factual correctness or license uniformity.

Biases, Risks, and Limitations

• Repository popularity, public availability, upstream sampling, and filtering choices affect which languages, communities, and software practices are represented.
• Examples may contain vulnerabilities, obsolete APIs, incorrect answers, toxic language, personal data, or other undesirable upstream artifacts.
• Exact duplicates are removed, but near-duplicates and benchmark contamination may remain.
• Language, license, timestamp, and provenance coverage varies by source.
• The buffered streaming shuffle is deterministic for the recorded configuration but is not a global random permutation.
• Several upstream revisions are unpinned, so the metadata may not be sufficient to recreate identical upstream streams in the future.

Licensing Information

This is a mixed-license dataset. Applicable terms vary by example and may include permissive software licenses, public-domain material, open-content licenses, CC BY-SA, and other licenses present in mixed-code sources. Per-record license and provenance information is retained in license, url, repository, and metadata.

Users are responsible for reviewing and complying with the applicable terms, including attribution and share-alike requirements. License labels are inherited from upstream metadata and have not been independently verified. This dataset card is not legal advice.

Additional Information

Dataset Curators

Owen Qwen

Citation

@misc{qwen2026codev1,
  author    = {Owen Qwen},
  title     = {Code Corpus code-v1},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/owenqwenllmwine/code-v1}
}

Upstream Dataset Citations

All seven unique upstream dataset repositories are linked in Data Splits. The six common-pile datasets share the Common Pile citation. The Stack V2-derived code also uses The Stack V2 citation, while CommitPackFT uses the OctoPack citation.

@article{kandpal2025common,
  title   = {The Common Pile v0.1: An 8TB Dataset of Public Domain and Openly Licensed Text},
  author  = {Nikhil Kandpal and others},
  journal = {arXiv preprint arXiv:2506.05209},
  year    = {2025},
  url     = {https://arxiv.org/abs/2506.05209}
}

@article{lozhkov2024starcoder2,
  title   = {StarCoder 2 and The Stack v2: The Next Generation},
  author  = {Anton Lozhkov and others},
  journal = {arXiv preprint arXiv:2402.19173},
  year    = {2024},
  url     = {https://arxiv.org/abs/2402.19173}
}

@article{muennighoff2023octopack,
  title   = {OctoPack: Instruction Tuning Code Large Language Models},
  author  = {Niklas Muennighoff and Qian Liu and Armel Zebaze and Qinkai Zheng and
             Binyuan Hui and Terry Yue Zhuo and Swayam Singh and Xiangru Tang and
             Leandro von Werra and Shayne Longpre},
  journal = {arXiv preprint arXiv:2308.07124},
  year    = {2023},
  url     = {https://arxiv.org/abs/2308.07124}
}

These scholarly citations do not replace the attribution, notice, or share-alike requirements of the licenses attached to individual records.

Reproducibility

manifest.json records the build timestamp, requested and realized token counts, document and filter statistics, upstream revisions, tokenizer, and shard inventory. recipe.json contains the complete source weights, filters, and processing configuration for the release.