File size: 835 Bytes
32f387d
 
 
 
 
 
 
1
2
3
4
5
6
7
8
# Anholon Phase 1 continuation mathematics

This frozen training subset contributes 200,000,000 loss-bearing mathematics labels to the 4B continuation. It uses the Phase 1 SentencePiece tokenizer and 2,048-position NumPy arrays. The same archive and manifest are available in both Anholon-Pretrain and Cayley-Math. Cayley contributors maintain their own collections in `components/`.

Extract `math-000.tar.zst` with `tar --zstd -xf` and place a copy of `manifest.json` at the extraction root. The manifest names the frozen array paths and their digests. `archives.json` records the archive digest and its parent training-manifest digest.

The [attribution notes](ATTRIBUTION.md) identify upstream works, applicable licenses, and transformations. The [source ledger](math-attribution.jsonl.gz) maps the attribution to the packed rows.