Attribution

Lux adapts the text backbone and tokenizer of Qwen3.5-9B at the pinned source revision, developed by the Qwen team at Alibaba Cloud and released under Apache License 2.0. The original Qwen license is retained without modification. Lux excludes the vision tower and vocabulary-generation head, adds a shared candidate head, and adapts the text model for typed decisions.

The retained decision mixture includes programmatically verified tasks, human-labeled evidence judgments from MultiNLI and replay from BANKING77 and CLINC150. It is a subset of the established training pool; Lux does not inherit the trained Sol or Nox weights. Official Jev outputs are not used as training labels.

BANKING77, by Iñigo Casanueva and colleagues, from the PolyAI source repository, CC BY 4.0. Original intent annotations are converted to runtime-defined choices.
CLINC150, by Stefan Larson and colleagues, from the CLINC source repository, CC BY 3.0. Source domain and intent labels are retained in the decision adaptation.
MultiNLI, by Adina Williams, Nikita Nangia and Samuel R. Bowman, NAACL 2018. The retained subset uses the non-fiction government, slate, telephone and travel genres; fiction is excluded. The pinned dataset card describes the source-specific terms, including the Open American National Corpus's permissive terms.

Natural-language decision adaptation

Lux uses 8,000 human-annotated training examples alongside 16,000 retained decision examples: 4,000 Cosmos QA reading questions, 2,000 SQuAD 2.0 answerability judgments, and 2,000 SNLI inference pairs. The latest adaptation stage completes one additional pass of this 24,000-example mixture. Human source labels are preserved; SQuAD answerability uses its supplied impossible/answerable annotation, and each source is converted to the model's decision interface. Source-parent groups and near duplicates are separated between custom training, selection and calibration. No official Jev output supplies a training label.

Cosmos QA, by Lifu Huang, Ronan Le Bras, Chandra Bhagavatula and Yejin Choi. Data from the author repository at the pinned revision. The official AllenAI dataset card records the author-confirmed CC BY 4.0 license.
SQuAD 2.0, by Pranav Rajpurkar, Robin Jia and Percy Liang. The official SQuAD project distributes the dataset under CC BY-SA 4.0. The adaptation here is answerability classification, not the official span-extraction benchmark.
SNLI 1.0, by Samuel R. Bowman, Gabor Angeli, Christopher Potts and Christopher D. Manning. The official Stanford Natural Language Inference project and release README identify CC BY-SA 4.0 for the corpus. Original entailment, neutral and contradiction labels supply the three decision alternatives.

QuALITY, by Richard Yuanzhe Pang and colleagues, contributes separate article-disjoint selection and calibration partitions and an additional held-out reading diagnostic. The released model did not use the alternative QuALITY training-source replacement. The author repository at the pinned revision distributes the dataset under CC BY 4.0; individual source-passage terms remain applicable.

These dataset licenses govern their respective source material; they are not replaced by the model package's Apache 2.0 license. Source corpus text, transformed training records and individual evaluation predictions are not redistributed in this model package. New evaluation panels, including supplied-fact QASC questions, are described separately in EVALUATION.md; evaluation data are not used for checkpoint selection or temperature fitting. As with other public datasets, exclusion from this custom training does not establish absence from upstream pretraining.