Commercial model training clarification for HotpotQA — Caedral

[sup...@caedral.com]

Caedral AI infrastructure and retrieval systems Hello HotpotQA Team, My name is Leonardo Turque, founder of Caedral. We develop AI infrastructure and CPU-efficient embedding and reranking systems. We are evaluating HotpotQA as a training source for proprietary English retrieval models. We understand that the dataset and the processed Wikipedia corpus are distributed under Creative Commons Attribution-ShareAlike 4.0. We would like written clarification regarding commercial model training, private model weights and hosted inference. Requested clarification Whether HotpotQA may be used to train proprietary commercial embedding and reranking models. Whether resulting model weights may remain private and be served commercially through hosted APIs. Whether filtering, deduplication, transformation, hard-negative generation, teacher scoring and combination with other datasets are permitted. Whether ShareAlike applies to model weights and internal training artifacts, or only to redistributed adaptations of the dataset. Which attribution language should appear in model cards, dataset cards and public documentation. Caedral would not redistribute the raw dataset. Access would be restricted to the training environment, with provenance, versioning, security and deletion controls. Could you please confirm whether the uses above are already permitted by the current license, or direct us to the appropriate rights holder for a separate written authorization? Thank you for your time. Leonardo Turque Founder, Caedral https://caedral.com sup...@caedral.com @trycaedral Caedral · AI infrastructure, embeddings and reranking https://caedral.com