Datasets:
Composition, measured
Numbers below are text bytes after cleaning, not file sizes (JSONL adds
~10 % of metadata). The final mix is produced by build, which applies
deduplication and the quotas; the "fetched" column is the pool it draws from.
Fetched pool
| EN | RU | total | |
|---|---|---|---|
| text | 4.9 GB | 2.4 GB | 7.3 GB |
| documents | 690k | 105k | 795k |
At the measured 3.34 bytes/token on the 8K bilingual BPE this is ≈ 2.15B tokens before dedup and quota selection.
By domain
| domain | fetched share | notes |
|---|---|---|
| mathematics | 26 % | OpenWebMath is the bulk; Wikipedia math adds definitions |
| science | 22 % | arXiv full text dominates; cosmology curated separately |
| encyclopedic | 16 % | Wikipedia EN/RU |
| code | 14 % | CodeParrot Clean + CPython stdlib |
| literature | 11 % | Gutenberg novels, Russian classics from Wikisource |
| ML | 5 % | arXiv abstracts, textbooks, framework docs |
| logic | 2 % | Lean mathlib proofs + Wikipedia logic |
| CS / engineering | 3 % | curated Wikipedia categories (chips, VLSI, XP, agile) |
| humour | <1 % | Gutenberg satire, transcripts |
| identity | <0.1 % | synthetic, 800 plain + 850 pastiche per language, no numeric facts |
Language ratio
60/40 by bytes is the contract. Because Cyrillic costs ~1.8× more bytes per
character, the same split is ≈ 55/45 by bytes but only ≈ 33 % Russian by
tokens with the 8K BPE. Both numbers are reported in tokenizer_meta.json;
the byte contract is the stable one, the token share depends on the tokenizer.
Shortfalls (kept visible)
- logic — 10 % quota, ~2 % realised. Formal proofs exist (mathlib), but general logic prose is thin.
- ML — 12 % quota, ~5 % realised. arXiv abstracts are high quality but short; textbooks are paywalled.
- humour — 1 % quota, ~1.5 MB realised. Public-domain satire and dialogue are small by nature.
- Russian code — the code quota is filled with English source; Russian code comments are not a coherent corpus of their own.
The mixer emits these as shortfalls in corpus_report.json rather than
padding the quota with mislabelled text.
Dedup and held-out
- MinHash, 64 permutations, 5-shingles, Jaccard threshold 0.8.
- Held-out: 2 % of each (language, domain) bucket, at most one document boundary of overshoot.
- Buckets under 5 MB (identity, rare slices) are not split, so they stay in training in full.