[gerardo-navarro]

Summary

Design proposal for discussion, no code yet. Organisations whose clients cannot reach Hugging Face (e.g. a city administration, see #318) want to put the models on an internal web server or network share once and have every aTrain installation load them from there.

Proposal in docs/proposals/internal-model-source.md:

• One setting model_source: a URL or a path (incl. UNC). When set, aTrain loads models only from there and never contacts Hugging Face.
• Layout at the source = a normal local models folder, so IT can just copy one.
• download_model() gets a second branch that fetches every file listed in the models.json manifest, then the existing verify_model() checks the pinned SHA-256. No Hugging Face API needed; a wrong or tampered source fails verification.
• Configuration: environment variable ATRAIN_MODEL_SOURCE first; later a config.toml (user, then machine-wide in %ProgramData% / /etc), read with stdlib tomllib.
• Also documents how aTrain is configured today (env vars only, no config file) and why HF_ENDPOINT alone does not fit a plain web server or share.

Feedback wanted on the open questions at the end of the doc, especially: no fallback to Hugging Face when a source is set.

Note: the doc links to docs/offline-model-install.md, which #320 adds.

References

Refs #318, related to #320

Screenshots / Examples

n/a (proposal)

Verification Steps

• Read docs/proposals/internal-model-source.md and comment on the approach and the open questions.
• Once agreed, a follow-up PR implements step 1 (ATRAIN_MODEL_SOURCE + download branch + tests + docs); this PR then either merges as the design record or is closed.

[is-hki]

Just want to leave a comment, that this could be very desirable in case of transcription of sensitive interviews which is quite commonplace in the academic field of humanities, for instance - handling of such data may be restricted to isolated environments by default.

[gerardo-navarro]

Just want to leave a comment, that this could be very desirable in case of transcription of sensitive interviews which is quite commonplace in the academic field of humanities, for instance - handling of such data may be restricted to isolated environments by default.

@is-hki Thanks for the feedback. I agree 💯 . We got this requirement while deploying aTrain in an enterprise IT environment which could be / is very similar to what you are describing.

We am working on a solution for this 👍

[JuergenFleiss]

Just want to leave a comment, that this could be very desirable in case of transcription of sensitive interviews which is quite commonplace in the academic field of humanities, for instance - handling of such data may be restricted to isolated environments by default.

Thanks for opening the discussion.

For default use I think that the current implementation of models being downloaded from a verified atrain huggingface repository is ok on a standard universtiy machine, as aTrain does not send any interview data anywhere. A number of universities have implemented atrain as is for local / offline interview transcription, including my own University of Graz.

Yes, but I would argue that this is only for the most critical data where you want no internet access from the machine anyway.

Do you know any cases where such a requirement was enacted?