storagestuffnobody/stack-v3.1-subsampled

Stack v3.1 subsampled

Extraction and upload are in progress. Available Parquet files are a partial dataset. Wait until manifest.json reports status: complete before using the complete subset.

A deterministic subset of HuggingFaceCode/stack-v3-train at revision 1f61b735bc0a5698345ce2196730f24bfa467f33. Every qualifying repository row is retained.

The earlier sample projected about 44.2 billion estimated tokens. Final counts will replace this estimate when extraction and verification finish.

Selection

All conditions apply after removing screened files:

• Estimated repository length between 16,384 and 1,048,576, inclusive.
• GitHub stars greater than 10 (at least 11), using the source snapshot's metadata.
• At least one retained file matching the test or CI path heuristics.
• At least four retained source files.
• Source files account for at least 50% of estimated repository tokens.
• No single retained file accounts for more than 50% of estimated repository tokens.

The path filter removes vendor/dependency directories, selected generated/build paths, minified JavaScript/CSS, source maps, lockfiles and generated protobuf filenames. Root support files and CI files are exempt from the vendor flag exclusion; explicit excluded paths still apply. Exact regular expressions are published in file_flags. Source-language classification and all thresholds are in selection_config.json.

Token estimates sum each retained file's upstream size_bytes multiplied by its language's published aggregate token/byte ratio, from the source dataset's Qwen/Qwen3-Coder-Next-Base statistics. No tokenizer was run on these repositories. Serialization, separators, paths and training wrappers are not counted; actual model-token lengths can fall outside this estimated range.

Format and use

Each row preserves repo_path, repo_id, commit_id, github_metadata and the retained files list. num_files is updated. Original file order, text and file metadata, including license fields, are retained. estimated_tokens is a floating-point proxy count. selection_metadata records input file/token counts, source composition, largest-file and removed-token fractions, test/CI/manifest flags, and the original shard/row.

from datasets import load_dataset

data = load_dataset("storagestuffnobody/stack-v3.1-subsampled", split="train", streaming=True)
repository = next(iter(data))
for file in repository["files"]:
    print(file["file_path"], file["content"][:100])

The Parquet files can also be read with PyArrow. statistics.json contains language and length distributions. manifest.json contains every Parquet file's SHA256, size and row count. source_manifest.json identifies all pinned upstream shards and their Hub hashes.

To reproduce, download that upstream revision, then run the supplied helper with an empty work directory:

python reproduction/reproduce_stack_v3_subset.py --source-dir /path/to/source-snapshot --work-dir /path/to/new-work --workers 2

The extraction requires Python, NumPy, PyArrow and PyYAML. It scans several TB of compressed source data; choose worker and memory limits for available resources. The frozen extractor verifies eligibility and ordered content hashes after each Parquet write. The helper generates a local plan and a resumable command. Its optional --verify-source-hashes checks every upstream SHA256 before starting.

Interpretation and attribution

Stars and test/CI paths are heuristics, not measured code quality, successful builds or passing tests. No extra content classifier, family/owner deduplication, language quota, license gate, dependency validation, training-window construction or evaluation holdout is added. Removing files may leave incomplete dependencies. This subset has not been evaluated for downstream model quality or benchmark contamination.

This dataset derives from The Stack v3 by Hugging Face / BigCode. See the source dataset card for collection, upstream filtering, licensing and removal information, and Am I in The Stack? for upstream removal requests. Database attribution follows ODC-By 1.0; individual files retain their original license metadata. The included GitHub Linguist taxonomy retains its MIT license.