AIIT-Tessera24B-dataset

The ~24.5B-token pretraining corpus used to train Tessera 1B, from AIIT-THRESHOLD.

What makes it different: this is not a single lab dump. The bulk is web text (DCLM), but the long tail was hand-selected, source by source — a deliberate curation rather than a firehose.

Contents

Tokens	24,504,827,904 seen (~24.5B); curated set ≈24.66B
Shards	674, tokenized with the Tessera tokenizer (byte-level BPE, vocab 65,536, memory-organ atoms)
Format	uint16 token id streams (.u16.bin) + tokenizer + shard manifest (sha256)
Languages	English (primary), Japanese

Sources

Bulk: DCLM (web). Hand-selected long tail: Wikipedia, Project Gutenberg, arXiv, PubMed Open Access, USPTO patents, mC4-ja, SlimPajama, RefinedWeb, The Pile, Dolmino, HackerNews, newspapers, philosophy texts, sacred texts, NSF award abstracts, OSF, CrossRef open-access abstracts, DOAJ, GitHub issues, Magicoder-OSS-Instruct, Cosmopedia-v2, Kyoto-school texts.

Data policy & honesty notes

• No model-conversation transcripts. No synthetic reasoning traces. (Per AIIT's training-data policy.)
• Two components are third-party synthetic: Cosmopedia-v2 (Mixtral-generated) and Magicoder-OSS-Instruct. Disclosed, small, public datasets — not our own model outputs.
• Deduplication is exact-match only. Fuzzy/near-dedup (MinHash) did not complete; expect some near-duplicate content.

⚠️ Redistribution — read before download

This corpus is derived from many upstream datasets with different, sometimes restrictive, licenses. AIIT does not hold redistribution rights to all of them. Therefore this release is structured in two tiers:

1. Tokens published directly — shards from clearly redistributable sources (e.g. Wikipedia CC-BY-SA, Project Gutenberg public domain, arXiv, PubMed OA, USPTO patents, NSF/OSF/CrossRef/DOAJ open access, and permissively-licensed web).
2. Recipe only (not the tokens) — for sources whose licenses do not permit redistribution (e.g. The Pile subsets, OpenSubtitles, copyrighted newspapers). For these we publish the exact source list, filters, and tokenizer so you can reproduce those shards from the public originals yourself.

This keeps the release genuinely open without stepping on upstream terms. Every shard's tier is marked in MANIFEST.md.

License

Per-source (see SOURCES.md). AIIT's own curation metadata, manifest, and the tessera_tokenizer.json are released Apache-2.0.