# Hummingbird-V1 training-data notices

This file documents the data lineage for the released `Hummingbird-V1` weights.
Apache-2.0 applies to the project code and released model materials; it does not
replace the licenses or notices attached to third-party source datasets.

## Training lineage

The selected checkpoint has 2,000,170,752 cumulative token presentations:

- 1,500,000,000 presentations in the original Muon causal-pretraining phase;
- 500,170,752 presentations in the natural-corpus continuation phase.

The continuation used a 1,091,660,174-token packed training split. Its complete
prepared-corpus totals across train, validation and held-out splits were:

| Source | Frozen revision | Prepared tokens | License |
|---|---|---:|---|
| FineWeb-Edu (`sample-100BT`) | `87f09149ef4734204d70ed1d046ddc9ca3f2b8f9` | 570,339,212 | [ODC-By 1.0](https://opendatacommons.org/licenses/by/1-0/) |
| DCLM baseline 1.0 | `a3b142c183aebe5af344955ae20836eb34dcf69b` | 211,970,663 | [CC-BY 4.0](https://creativecommons.org/licenses/by/4.0/) |
| FineWeb-HQ | `e58199cdd52438d94405df1a4d8630cc5f13bf84` | 109,226,534 | [ODC-By 1.0](https://opendatacommons.org/licenses/by/1-0/) |
| SmolLM-Corpus / Cosmopedia v2 | `3ba9d605774198c5868892d7a8deda78031a781f` | 168,663,969 | [ODC-By 1.0](https://opendatacommons.org/licenses/by/1-0/) |
| FineMath (`finemath-4plus`) | `e92b25a616738fe95dc186b64dfb19f9c8525594` | 53,640,449 | [ODC-By 1.0](https://opendatacommons.org/licenses/by/1-0/) |

The original base phase also used FineWeb-Edu, Cosmopedia v2, TinyStories, and
project-generated procedural arithmetic, tutorial and short-choice-rationale
text. TinyStories revision `f54c09fd23315a6f9c86f9dc80f725de7d8f9c64`
is made available under CDLA-Sharing-1.0.

## Filtering, deduplication and evaluation protection

The natural corpus used quality admission filters, canonical exact-document
deduplication, and 32-permutation MinHash/8-band LSH near-duplicate removal
confirmed by exact Jaccard similarity at a 0.80 threshold. Documents were split
deterministically by SHA-256 before packing.

Rendered Open SLM and ArithMark-3 prompts plus choices were held in a protection
index and excluded from training. Public benchmark results were nevertheless
evaluated at the 250M, 500M, 750M and 1B continuation checkpoints and used to
select the released 500M checkpoint. This checkpoint-selection bias is disclosed
in the model card; benchmark records and answers were not training data.

Exact quotas, filtering, source fields, split policy and prepared token counts
are included in `training/corpus_contract.yaml`, `training/packed_metadata.json`
and `training/provenance.json` in the release package.