dlab-spp/corpus-1T-manifest

SPP Corpus 1T Manifest

The selection manifest for the ~1.0T-token pretraining corpus used in Synthetic Persona Pretraining (SPP): Alignment from Token Zero.

The corpus is a seeded subsample of allenai/dolma3_mix-6T. Rather than redistribute ~2.6 TB of text that is already public, this dataset publishes the selection decisions keyed by upstream document id, so the corpus can be reconstructed exactly by replaying against upstream.

📄 Reflections + text for the annotated half: dlab-spp/reflection-50m

🧾 Safety classifier outputs: dlab-spp/safety-classifications

Configs

config	rows	what
annotated	102,772,028	rows selected for reflection annotation (safety_score >= 3 plus a matched sample)
unannotated	925,065,551	the rest of the 1T corpus
text_rescue	27,904,353	text for rows with no upstream id (see below)

Columns in annotated / unannotated: id, safety_score, has_annotation, is_bad. Total 1,027,837,579 rows = 1.000T tokens.

Reconstruction

1. Fetch text from allenai/dolma3_mix-6T at revision 689a3ea2d8217e64d73a5058913fa43ad15e81aa (pin this — the shard manifest is a seeded shuffle over the shard count, so it changes if upstream does).
2. Join by id.
3. Splice in text_rescue rows at their recorded global_row positions.
4. Keep the row order exactly as published — the Megatron .bin layout, and therefore batch composition, depends on it.

file_boundaries.json maps each of the original 40,000 subsample part files to [global_start_row, n_rows], so the original partitioning is reconstructible without publishing 40,000 files. subsample_metadata.json carries the seeded subsample parameters verbatim (seed=42, annotation_threshold=3, scale=0.871747139193521).

Full procedure, including tokenization and the reflection insertion, is in REPRODUCTION.md in the code release.

Why text_rescue exists

stack_edu-Python is the only 1 of the 163 subsets in allenai/dolma3_mix-6T that was never normalized to the Dolma schema. Its 301 shards ship raw Stack-Edu fields flattened at top level — blob_id, repo_name, path, score, … — with no id field at all, unlike its 14 sibling language subsets. Rows drawn from it therefore have a null id and cannot be joined back to upstream.

split	rows without id	share
annotated	1,629,168	1.585%
unannotated	26,275,185	2.840%
total	27,904,353	2.715%

Rather than leave a 2.7% hole, the text_rescue config publishes those rows' text directly, keyed by global_row. Merge them at their recorded positions and reconstruction is complete.

(Upstream does carry usable identifiers for these rows — blob_id, or repo_name + path. They are simply not named id and were not among the columns downloaded.)

License and attribution

Released under the Open Data Commons Attribution License (ODC-BY 1.0), inherited from the upstream source.

Contains information from allenai/dolma3_mix-6T, made available under the Open Data Commons Attribution License (ODC-BY 1.0).

Please cite Olmo 3 (arXiv:2512.13961) and observe AI2's Responsible Use Guidelines. Upstream frames this data as intended for research and educational use; that framing carries over here.

Citation

@misc{minder2026syntheticpersonapretrainingalignment,
      title={Synthetic Persona Pretraining: Alignment from Token Zero},
      author={Julian Minder and Viktor Moskvoretskii and Raghav Singhal and Difan Jiao and Andy Arditi and Shaobo Cui and Yiderigun Borjigin and Kartik Bali and Stefan Krsteski and Harsh Raj and Huu Nguyen and Jannik Brinkmann and Ashton Anderson and Roland Aydin and Robert West},
      year={2026},
      eprint={2608.13482},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2608.13482},
}