sunnysetia's picture
normalize tokenized pool shard layout and organization linkage
f6251f5 verified
|
Raw History Blame Contribute Delete
932 Bytes
metadata
license: other
task_categories:
  - text-generation
language:
  - en
pretty_name: Locus EXP Commit Code (Dolma 2 tokenized)

Locus EXP Commit Code - Dolma 2 tokenized

Pretokenized experiment pool for reproducible proxy-training runs.

  • MANIFEST.json is the authoritative schema, provenance, checksums, and train/holdout assignment.
  • shards/shard-NNNNN/tokens.bin stores little-endian int32 token IDs.
  • offsets.bin stores little-endian int64 document boundaries.
  • index.parquet stores document IDs, offsets, and compact filter fields.
  • metadata.parquet stores complete source metadata and is downloaded only for explicit cold-metadata filters.

Each document already ends in exactly one EOD token. Training must not add another document-level EOD.

Commit examples use target_start_token to mask the commit-message prompt and train only on the unified-diff target. Each commit document must fit one sequence.