open-athena/snowball-upstream-ids

Snowball upstream IDs

This dataset does the document-ID-to-upstream-ID join for the Snowball replay index. It contains no source text or token arrays. Each Parquet file has one nullable UTF-8 column, upstream_id, and is row-for-row aligned with the same-numbered data/selection file in open-athena/snowball-replay-index revision 122e87d33518cbe4587dcaba59662d09bbdf27a0.

For example, row 17 in data/upstream_ids/part-00042-of-06301.parquet describes row 17 in data/selection/part-00042-of-06301.parquet. No hash join or Marin installation is required:

import pyarrow.parquet as pq
from huggingface_hub import hf_hub_download

name = "part-00042-of-06301.parquet"
selection = pq.read_table(
    hf_hub_download(
        "open-athena/snowball-replay-index", f"data/selection/{name}", repo_type="dataset", revision="122e87d33518cbe4587dcaba59662d09bbdf27a0"
    )
)
upstream = pq.read_table(
    hf_hub_download("open-athena/snowball-upstream-ids", f"data/upstream_ids/{name}", repo_type="dataset")
)
assert selection.num_rows == upstream.num_rows
rows = selection.append_column("upstream_id", upstream.column("upstream_id"))

Join source_id from the selection row to sources.json in this repository. That record identifies the public upstream repository, revision, and the source field encoded in upstream_id. The sidecar preserves strings as UTF-8, integers as decimal strings, binary values as lowercase hexadecimal, and compound locators as canonical JSON.

Most sources expose a source-provided ID, including uuid for Nemotron-CC and Nemotron specialized data and nemotron_pretraining_id for Nemotron Code. ClimbLab-Ja, institutional books, and StarCoder documentation did not retain a unique row key; their entries contain the best surviving source or project locator and are marked non_unique_locator. Upstream IDs are otherwise source-provided identifiers, but this release does not claim that every upstream publisher enforces global uniqueness.

manifest.json records every sidecar's row count, byte count, SHA-256, and the SHA-256 of its paired selection shard. The extraction re-applied Snowball's decontamination and deduplication masks and required every retained (source_id, document_id) row to equal the paired replay-index shard before writing the sidecar.