# Training-source attribution

Lex inherits the following training-source lineage through its fixed Kai parent. This package redistributes model artifacts and project code, not the source datasets. Dataset licenses remain with their authors; the model's Apache-2.0 license does not grant new rights to those texts or annotations.

| Source | Attribution and source terms |
|---|---|
| SNLI | Bowman, Angeli, Potts and Manning (2015), [Stanford Natural Language Inference](https://nlp.stanford.edu/projects/snli/); CC-BY-SA-4.0. |
| BoolQ | Clark et al. (2019), [Boolean Questions](https://github.com/google-research-datasets/boolean-questions); CC-BY-SA-3.0, with underlying Wikipedia attribution retained in source records. |
| ARC | Clark et al. (2018), [AI2 Reasoning Challenge](https://huggingface.co/datasets/allenai/ai2_arc); publisher metadata identifies CC-BY-SA-4.0. |
| MASSIVE | FitzGerald et al. (2022), [Amazon MASSIVE](https://github.com/alexa/massive), multilingual intent labels. |
| HelpSteer2 | Wang et al. (2024), [NVIDIA HelpSteer2](https://huggingface.co/datasets/nvidia/HelpSteer2), human ordinal response ratings. |
| KLUE-STS | Park et al. (2021), [KLUE](https://github.com/KLUE-benchmark/KLUE); retain the dataset's CC-BY-SA-4.0 terms. |
| JSTS and JCommonsenseQA | Kurihara, Kawahara and Shibata (2022), [JGLUE](https://github.com/yahoojapan/JGLUE/tree/6f071c09316baae89c3d083a90985b4b1cb9968c); CC-BY-SA-4.0. |
| Natural Instructions | Wang et al. (2022), [Super-NaturalInstructions](https://github.com/allenai/natural-instructions/tree/55a365637381ce7f3748fa2eac7aef1a113bbb82). Instruction definitions and dataset instances have distinct terms; the collection's code license does not replace instance-source licenses. |
| SocialIQA | Sap et al. (2019), [Social IQa](https://allenai.org/data/socialiqa); CC-BY-4.0. |
| QASC | Khot, Clark, Guerquin, Jansen and Sabharwal (2020), [QASC](https://github.com/allenai/qasc/blob/c44a320c2bb682d614a675d2ba4be072150db4eb/README.md); CC-BY-4.0. |
| Cosmos QA | Huang, Le Bras, Bhagavatula and Choi (2019), [Cosmos QA](https://github.com/wilburOne/cosmosqa/tree/b6eb99cca4e2a51dd28a9a6f562534872d851639); [official publisher card](https://huggingface.co/datasets/allenai/cosmos_qa) records CC-BY-4.0 and author confirmation. |
| TyDi QA | Clark et al. (2020), [TyDi QA](https://github.com/google-research-datasets/tydiqa); Apache-2.0 dataset/repository terms do not remove underlying Wikipedia attribution obligations. |
| OpenAssistant | Köpf et al. (2023), [OpenAssistant Conversations](https://huggingface.co/datasets/OpenAssistant/oasst1); Apache-2.0. Human quality/helpfulness annotations were used in training. |
| AfriSenti | Muhammad et al. (2023), [AfriSenti and SemEval-2023 Task 12](https://github.com/afrisenti-semeval/afrisent-semeval-2023/tree/5aec3cfcf87b59f885e98641f72e52182776429d); CC-BY-4.0. The upstream README also provides the NaijaSenti and ASAB citations. |
| SentiMix Hinglish | Patwa, Aguilar, Kar, Pandey, PYKL, Gambäck, Chakraborty, Solorio and Das (2020), [SemEval-2020 Task 9](https://aclanthology.org/2020.semeval-1.100/), [dataset DOI](https://doi.org/10.5281/zenodo.3974927); CC-BY-4.0. |

The Natural Instructions training families were query well-formedness (task 673), CaSiNo (353), ATOMIC (1216/1211), poem sentiment (833), HateXplain (1502/1503), WIQA (1548/1727), and ProPara (1568). Their recorded instance terms are respectively CC-BY-SA-4.0, CC-BY-4.0, CC-BY-4.0, CC-BY-4.0, MIT, Apache-2.0 and Apache-2.0. For WIQA and ProPara, this records the published NI instance declarations; it does not claim a separate re-audit of every original data archive. Qasper was a development family, not a training family.

Project transformations included explicit Choice/Noul/Score schemas, candidate identifiers, deterministic splits and component isolation, adjacent-level distributions for original continuous means, and one-hot representations of original hard labels. Sentiment categories remain coarse ordered sentiment, not fabricated continuous human quality ratings. TyDi uses the annotated supporting passage for Boolean prediction; it is not a retrieval task. OASST Spanish/Chinese rubrics and schema paraphrases were agent-authored changes, not new human annotations. Controlled rule/state examples were generated by the project. No original dataset is relicensed by these transformations.

Training-source attribution is distinct from evaluation-source attribution and from claims about what may have appeared in upstream encoder pretraining. Benchmark-only datasets and excluded candidates are not represented here as training sources.

## Lex specialization

Lex additionally uses the original TRAIN split of [LocalLLaMA/typed-decisions](https://huggingface.co/datasets/LocalLLaMA/typed-decisions), fixed revision `ea9306458d6e9563628369a3d1e72e362fb381d2`. The publisher identifies this English synthetic dataset as Apache-2.0. The 1,200 original cases contain 6,000 decisions across four workflows and 20 question schemas. The dataset's generated states and teacher-derived hard/soft targets measure agreement with that annotation process, not human ground truth or validated business outcomes.

The original TRAIN parquet SHA256 is `46a58d63edfd86e23229c78afe8b72307bb4ca9fb0e8df180cabb3c67ec9dcd5` (598,824 bytes). Lex preserves complete states, semantic candidates, question criteria, original hard labels, and explicitly normalized soft targets. The final training target is an equal mixture of the normalized soft distribution and the one-hot original hard label. All original TRAIN decisions appear exactly eight times. TEST was excluded from training and recipe selection. Dataset instances, gold labels, latent factors, and third-party model/SDK files are not included in this distribution.