# Training data attribution

Loom Spark 2 was trained on several openly licensed corpora. Some of these licences
require attribution; this file satisfies that requirement and must be kept with any
redistribution.

## SQuAD 2.0 — CC BY-SA 4.0
Rajpurkar, Jia & Liang. "Know What You Don't Know: Unanswerable Questions for SQuAD."
https://rajpurkar.github.io/SQuAD-explorer/
Used for grounded reading, and — via its unanswerable questions — for teaching the model
to say when a result does not contain the answer.

## MASSIVE — CC BY 4.0
Amazon. https://github.com/alexa/massive
Derived from SLURP, also CC BY 4.0. Used for tool-decision training.

## CLINC150 — CC BY 3.0
Larson et al. "An Evaluation Dataset for Intent Classification and Out-of-Scope
Prediction." https://github.com/clinc/oos-eval
Used for tool-decision training.

## databricks-dolly-15k — CC BY-SA 3.0
Databricks. https://huggingface.co/datasets/databricks/databricks-dolly-15k
Used for instruction following.

## OpenAssistant OASST1 — Apache 2.0
LAION / OpenAssistant. https://huggingface.co/datasets/OpenAssistant/oasst1
Used for multi-turn dialogue structure. Only English conversations with short assistant
replies were kept.

## Persona curriculum — Textile Labs
Identity, limits, warmth and brevity were written for Loom and are not derived from any
public dataset.

Only English portions were used. No source text was altered except for truncation of
passages to a realistic tool-result length, and surface augmentation (casing, punctuation,
filler) applied to user turns in training copies only.