| --- |
| pretty_name: YAT training data provenance and versions |
| language: |
| - multilingual |
| tags: |
| - provenance |
| - reproducibility |
| - yat |
| - mmbert |
| - metadata |
| --- |
| # YAT training data: provenance and versions |
|
|
| This public catalog records the distinct prepared corpora used or considered in FlaxChat YAT training. **This release contains metadata and references only. It does not yet contain raw training examples or token arrays.** The inventory is incomplete where historical artifacts have not been recovered. |
|
|
| Read [`catalog.json`](catalog.json) for status, upstream revisions, prepared sizes, metadata SHA-256 values, artifact references, and known gaps. `metadata/` holds the original recipe and manifest bytes. A `historical_trained` or active status is separate from candidate, development, and validation status. Prepared pool size is not training exposure. Recoveries reuse the same frozen corpus rather than defining new data versions. |
|
|
| ## Versioning |
|
|
| The immutable release tag is `catalog-2026-10-10`; consumers should pin its resolved Hugging Face commit. Each prepared dataset has a distinct ID and original content/recipe hashes. Corrections receive a new release tag; upstream revisions, splits, tokenizer identity, and sampling seeds must stay explicit. Future 10B preparation receives a separate version, with actual counts published only after completion. |
|
|
| ## Included families |
|
|
| * FineWeb2-HQ adaptation and pilot selections. |
| * Historical web/reference MLM mixtures (102M, 359M and 617M prepared nonpadding tokens). |
| * The current 1.043B multilingual/code MLM pool from October 6, reused by scratch GOAT-input training. |
| * Global Voices pilot, 28-language continuation, and larger carry-preserving continuation; the unused larger noncarry candidate is separate. |
| * MS MARCO, MIRACL and CodeSearchNet source references for the released embedding-v1 fine-tune. |
| * Representation-v2 development/repair references, explicitly incomplete. |
|
|
| ## Attribution and content licenses |
|
|
| Original sources remain credited in each recipe and upstream revision. FineWeb2-HQ, FineWeb2 and FineWeb-Edu declare ODC-BY dataset licensing; Common Crawl and underlying content terms still apply. Global Voices content is CC-BY-3.0: credit Global Voices contributors, OPUS alignment and Sentence Transformers packaging. The upstream package does not retain article URLs. GitHub code has per-file/repository licenses; a packaging license must not be treated as a blanket license for every code file. MS MARCO, MIRACL, CodeSearchNet and historical reference data require their own applicable source terms. No blanket license is asserted for this mixed catalog. |
|
|
| Only metadata is published here. Releasing content or token arrays requires preserving the corresponding provenance, attribution, and applicable notices; tokenization does not remove content obligations. GCS references document artifact identity and do not imply public access. |
|
|
| ## Quality limitations |
|
|
| The current 1B recipe performed normalized exact-document deduplication. It does not establish semantic or benchmark-wide decontamination. Historical corpora have different overlap policies recorded in their manifests. All counts and quality claims must be tied to the exact stage and revision; this catalog is not evidence of production quality or complete language/domain coverage. |
|
|
| Training implementation: [FlaxChat](https://github.com/mlnomadpy/flaxchat). Model family: [mlnomad on Hugging Face](https://huggingface.co/mlnomad). |
|
|
|
|
| ## Historical document reference release |
|
|
| The `references-2026-10-10` tag adds authenticated gzip JSONL document provenance for historical MLM pools. See [`references-index.json`](references-index.json) and [`references/README.md`](references/README.md). These files contain source coordinates and document metadata; **raw text and token arrays are not included**. The original `catalog-2026-10-10` tag remains unchanged. |
|
|
