Datasets:
Source attribution and modifications
The per-row ATTRIBUTION.jsonl maps exported conversation IDs to source identities. Original source metadata also remains in data/train.jsonl. Generated dialogue is a model-produced adaptation, not an original source quotation or a human-certified translation.
Wikipedia text
Source: Wikimedia Wikipedia dataset, credited to the contributors of each article identified by document_url in ATTRIBUTION.jsonl. Article links provide access to contributor histories. Source snapshot revisions and hashes are preserved; these are dataset revisions, not exact article revision IDs.
The pinned dataset card identifies original text under CC BY-SA 3.0 (and GFDL where applicable); retained source text keeps its original notice. Our adapted Wikipedia text is distributed under CC BY-SA 4.0, the later-version adaptation option in CC BY-SA 3.0 section 4(b). This notice applies to that material, not an assertion of exclusive ownership of the entire mixed package. See Wikimedia reuse terms, section 7. Retained source passages, references and notices are not stripped. Changes include excerpt selection, generated questions/answers, summaries, rewrites and multi-turn adaptation; there is no Wikimedia endorsement.
Synthetic task specifications
Math/code references and fictional tool scenarios are generated task specifications, not claims of real transactions. Odd-slot source-free multi-turn tasks are explicitly identified as synthetic. No external source identity is fabricated for these rows.
Audit limitations
Acceptance uses existing model reviews and deterministic checks. Empty-rationale technical recovery preserves the original complete keep verdict and does not invent a rationale. These are not human/native-language certification claims.