# Attribution

Loom Spark 3.2 was trained from scratch: random initialisation, then two short second passes
on its own weights. No third-party checkpoint was used. Grounded-reading rows embed real
encyclopedic prose. Some of these licences require attribution; this file satisfies that
requirement and must be kept with any redistribution.

| slice | source | licence |
|---|---|---|
| grounded reading + "the answer isn't in this passage" | SQuAD 2.0 (Rajpurkar, Jia & Liang) | CC BY-SA 4.0 |
| three-paragraph reading (same-article paragraphs) | SQuAD 2.0 | CC BY-SA 4.0 |
| multi-hop grounded reading | HotpotQA | CC BY-SA 4.0 |
| trivia reading (answer sentence from the entity's article) | TriviaQA | Apache 2.0 |
| when to reach for a tool | MASSIVE (Amazon) / CLINC150 (Larson et al.) | CC BY 4.0 / CC BY 3.0 |
| instruction following | databricks-dolly-15k | CC BY-SA 3.0 |
| multi-turn dialogue structure | OpenAssistant OASST1 | Apache 2.0 |
| encyclopedic passages | Wikipedia | CC BY-SA |
| identity, limits, warmth, memory within a chat, attribution, injection resistance | Textile Labs | MIT |

Only English portions were used. OASST conversations in which the original assistant named
itself were removed. Source text was not altered except for truncation to a realistic
tool-result length, and surface augmentation (casing, punctuation, filler) applied to user
turns in training copies only.

## Search queries — derived, not generated
Every `<lookup>` query in the training data is the subject of the question, derived
mechanically from the human-written sources above. No language model wrote any training
query or any training reply. No real user data was used. Model weights: MIT.