Licensing question: training data composition + commercial deployment of laya-multilingual

[JansonYan1]

Hi Nandakishor,
First — thank you for open-sourcing Laya. We are a small team in China building a product that checks whether an AI assistant's answer contradicts itself or the source text it was given. We run the 322M multilingual checkpoint locally and plan to fine-tune it on our own (self-produced) data, then ship it inside a paid subscription service.
Before we commit, we need four things on the record. We asked because we could not find them published anywhere:

1. Training data composition. The model card says only public, human-labelled datasets were used, with no LLM-generated labels — but the dataset list itself is not published. Can you share the list, or at least confirm whether it contains any dataset under non-commercial or research-only terms? We are specifically asking about XNLI (CC BY-NC 4.0) and RACE ("non-commercial research only"), since both appear in third-party write-ups of Laya. Even a yes/no per dataset helps.
2. Do the Apache-2.0 weights permit commercial deployment where we do not redistribute them? Our use is a hosted service: weights stay on our own servers, we never ship or resell the weights themselves. Our reading is that Apache-2.0 permits this with notice retention — please correct us if that is wrong.
3. Is there any research-use-only or other restriction not stated in the license field? Anything in the model card, a separate agreement, or an upstream encoder term that we would not see from the repo frontmatter.
4. Fine-tunes. If we fine-tune laya-multilingual on our own human-annotated Chinese NLI data and deploy that derived checkpoint commercially, is that permitted under the same terms? Any attribution requirement beyond the Apache-2.0 NOTICE?
   We are not asking for legal advice and will treat your answer as a factual statement, not a legal opinion — we have counsel review separately. We are happy to publish your reply (with credit) in our compliance notes so other teams do not have to ask the same question.
   Context if useful: we chose Laya over a Chinese alternative specifically because the alternative was fine-tuned on a CC BY-NC dataset, which rules it out for us. That is the whole reason this question matters.
   Thanks for your time.
   — janson

[jsubirat]

Hey @NandhaKishorM , could you please provide some feedback on the concerns raised by @JansonYan1 above? We are also considering on testing Laya for a hosted service, and would like to know the exact same points raised in this discussion. Thank you very much in advance.