Cayley-Math / phase1 /README.md
JunyiZou's picture
Mirror 821M Phase 1 mathematics labels and extend the full viewer
70ab9b4 verified
|
Raw History Blame
1.78 kB
# Anholon Phase 1 Mathematics Mirror
This component contains 50 million loss-bearing training tokens in 43,833 packed training windows. The development and test splits each have 1,024 labels. The mirror preserves the frozen mathematics windows and uses the bundled 32,768-piece SentencePiece tokenizer.
Extract `math/math.tar.zst` into the directory containing `math/manifest.json`. The [file manifest](upload-manifest.json) records the repository-relative archive hashes, and the math manifest records the extracted array hashes.
The [source catalog](sources.json) and [attribution guide](attribution/README.md) identify the upstream works and their licenses. The preparation selected and normalized text, encoded the selected passages with the Phase 1 tokenizer, and packed the token sequences into causal windows.
### P1 mathematics excerpts from the Stacks Project
Original work: *The Stacks Project*, Johan de Jong and the Stacks Project contributors. Copyright © 2005–2025 Johan de Jong. Original source and history at [the fixed revision](https://github.com/stacks/stacks-project/tree/a04446e57ec1fbc252a871afcec7752fb2807b14). The existing copyright, contributor list, and license are in `attribution/math-rights/`. The original uses the GNU Free Documentation License version 1.2 or later, with no Invariant Sections or Cover Texts. This 2026 modification, published by MMLA-ORG, selects passages, omits presentation markup, and encodes the remaining text and math in P1 windows. The modified excerpts retain the same license terms.
The [first continuation](math/continuation-4b/README.md) adds 200M labels. The [full pretraining math continuation](math/pretrain-rest-16p42b/README.md) adds another 821M, for 1.071B training labels across the three frozen batches.