blab-jhu/KYS-3B-pretraining-data

KYS 3B pretraining data — four settings, seed 42

Training data for the KYS 3B grid: a 57-layer, 2,994,444,288-parameter llama model (the 1.5B architecture with only the depth changed), trained for 20B / 40B / 60B tokens (WSD; finals at steps 9537 / 19073 / 28610). Code: https://github.com/imHuicongZhang/nanotron/tree/release/kys-3b-600m at cd4ba52ab83807bc7a9502046a846829249849c0. Trained models will be published to blab-jhu/KYS-3B-pretraining-models.

setting	parts	rows	train tokens (assembled)	empty-text docs	tokens (tokenized)	documents (tokenized)
quality-first	65	32,341,160	20,000,000,270	27	20,000,000,243	32,341,133
diversity-oriented	54	26,692,847	19,834,314,433	2	19,834,314,431	26,692,845
disagreement-aware	54	26,743,280	20,000,000,227	8	20,000,000,219	26,743,272
quality-base	36	17,942,518	20,000,000,299	0	20,000,000,299	17,942,518

• quality-first / diversity-oriented / disagreement-aware: the shared raw base (10,527,887 docs, 10,000,000,186 train tokens: top DCLM fastText) plus a rewritten 10B strategy block.
• quality-base: the same shared base plus the NEXT 7,414,631 documents (10,000,000,113 train tokens) in the same fastText order, raw, so the corpus is the global fastText top ~20B prefix with no overlap or duplicate.
• Train tokens are llama-2 tokens + 1 EOS per document. Datatrove skips documents with empty text, so the tokenized stream is the assembled count minus the empty-text documents (rows left empty by the rewrite strip). diversity-oriented is 165,685,567 tokens short of 20B (accepted; its epochs are re-read accordingly, reports/tokenize/).

Layout

• corpora/<setting>/part_*.parquet: the final shuffled corpus (seed 42), in training order, unchanged.
• tokenized/<setting>/NNNNN_unshuffled.ds{,.index,.metadata}: nanotron/datatrove shards, shard i = part i. After downloading run kys3b.py relocate-metadata (line 1 of each .metadata records the path the shards were written under; only that text line changes, the .ds / .index binaries do not).
• tokenizer/: the exact 1.5B llama-2 tokenizer (sha256 in HANDOFF_RELEASE.json).
• reports/: assembly/selection validation, tokenization statistics, LR-resume regression, GRID_MANIFEST.json.
• RELEASE_MANIFEST.json: size and sha256 of every file, ordered parts and shards with rows/tokens/documents.
• HANDOFF_RELEASE.json: code commit, grid, model, tokenizer and initialization provenance, totals.
• receipts/UPLOAD_RECEIPT.json: the verified content revision and the verification results.

No initialization checkpoint is included: none exists yet. The operator creates it once on GPU (deploy/portable_3b/create_init.sh), and all four settings start from it.

Training with this data

Use deploy/portable_3b/ of the code branch: README.md there lists the install, the operator config (operator.template.yaml: your paths, environment, W&B entity/project, deployment and scheduler; nothing is defaulted) and the command sequence. kys3b.py check --corpus verifies a download against RELEASE_MANIFEST.json (--hash also hashes every token binary).

Provenance and licensing

This release does not set its own license (license: other). The terms of the inputs below apply. They have not been reviewed for this release, so check each upstream source before reusing the data.

• Source text: a 600M-document sample of DCLM-RefinedWeb (DataComp-LM), web text derived from Common Crawl. All raw documents (the shared base and the whole quality-base block) are the source text unchanged (checked by the stage-04 validation, reports/stage04_assembly/).
• Rewritten text (the 10B strategy block of quality-first / diversity-oriented / disagreement-aware): generated by Qwen/Qwen2.5-7B-Instruct (vLLM, bf16) from those DCLM documents, with the 1.5B KYS prompt set byte for byte.
• Selection signals: the DCLM fastText quality ranking and the KYS ModernBERT quality scorer (blab-jhu/KYS-Modernbert-Quality-Scorer@0afde2e6).
• Tokenizer (tokenizer/): the Llama-2 tokenizer files (via unsloth/llama-2-7b). Llama-2 materials are distributed under Meta's Llama 2 Community License.
• Selection and assembly evidence, including a recorded -dirty selection-code tag and its reproduction check: reports/stage04_assembly/PROVENANCE_SELECTION.md.