open-athena/snowball-replay

Snowball Hugging Face row replay

This repository identifies the upstream Hugging Face rows retained in the Snowball pretraining store. It contains row locators and mixture metadata, not source documents or token arrays. A reader does not need Marin, private GCS access, or either of the earlier Snowball index repositories.

Get the selected rows

Install Python 3.12 or newer, then run:

python -m pip install huggingface_hub pyarrow zstandard
hf download open-athena/snowball-replay README.md REFERENCE.md sources.json sources.parquet mix.json materialize.py \
  --local-dir snowball-replay
cd snowball-replay
python materialize.py --source-id 15 --output-dir snowball-source-15

materialize.py downloads the pinned upstream files needed for that source and writes the selected records locally. Parquet inputs remain Parquet; line-oriented JSON inputs become JSONL. Each record includes snowball_bucket_id and snowball_replay_order. Use sources.json or sources.parquet to choose a source and see its repository, revision, document count, token bounds, and access requirements.

The locator files are partitioned as data/source_id=NNNNN/part-*.parquet and have this schema:

Column	Type	Meaning
data_file	string	Path in the pinned Hugging Face dataset repository
row_index	uint64	Zero-based physical row in that file
bucket_id	uint8	domain_cluster * 5 + quality_bucket
replay_order	uint64	Position in the retained Snowball store before bucket filtering

For a custom reader, group locators by data_file, download that file from the repository and revision in sources.parquet, and select the listed physical rows. Sort on replay_order when store order matters. The source repository's license and access terms continue to apply.

Coverage and limits

The manifest covers 95 sources with an immutable Hugging Face revision, an unchanged historical file layout, no source-specific transform, and a unique upstream identifier. Those sources contain 11,193,046,614 of 11,728,558,302 retained documents. Their attributed token mass is between 9,376,772,787,706 and 9,632,173,243,255 of 10,372,343,704,053 tokens, or approximately 90.4% to 92.9%. Token bounds are necessary because 13.3 billion tokens occurred in store-task boundary shards that contained more than one logical source.

The remaining sources are recorded in the source table with an exclusion reason. They include transformed inputs, sources without a public Hugging Face origin, and three sources whose upstream identifier was not unique. The largest exclusion is ClimbLab-Ja, about 3.4% of token mass. Reproducing it would require a content-hash join rather than a direct row-ID lookup.

Within mapped sources, an upstream identifier can occasionally name repeated rows with identical normalized content. Those rows are interchangeable for document membership but do not have a uniquely recoverable historical occurrence. The release chooses the first matching physical row in lexical file and row order and never emits the same physical row twice.

This is a document-membership replay, not proof of the exact tokens observed by every training step. Exact token exposure also depends on the pinned tokenizer, document normalization, EOS handling, dataset shuffle, mixture block rounding, restarts, checkpoint sampler state, and the 65,536-token cooldown. mix.json preserves the recovered sampler configuration; REFERENCE.md explains the reconstruction and validation in detail.

Upstream repositories can disappear, change access policy, or require acceptance of separate terms. inventory-validation.json records the pinned-revision file check. The pinned revisions, file paths, and byte lengths were validated when this artifact was produced. Byte-length equality is not a cryptographic content proof; the mapping also relies on the historical artifacts' recorded download lineage. The upstream data itself is not mirrored here.