Clarification on externally hosted training-time artifacts

[KenMalloy]

Trying to get a clearer read on where the reusable artifact boundary actually sits.
I've seen public examples of custom tokenizers and pre-tokenized FineWeb shards used as reproducible setup artifacts: SentencePiece packages, Hugging Face-hosted retokenized datasets, etc. I get that framing: it's preprocessing infrastructure, the tokenizer choice affects BPB accounting, and the provenance is auditable. Fine.

What I'm less sure about is how far that precedent extends.

If a submission pulls down a pinned, public artifact at setup/training time, something derived from the allowed corpus, not used as a lookup during eval, not loaded as extra runtime state by the final model, but it does influence training (e.g., as initialization, optimizer state, auxiliary statistics, a small learned prior), is that within bounds? My working assumption is:

• tokenizer and pre-tokenized data artifacts are acceptable as reproducible preprocessing

• learned weights, checkpoints, or auxiliary models produced outside the timed run are a different category: they either need to fit in the 16 MB submitted artifact, be regenerated inside the timed run, or be left out entirely

Do people feel that reading roughly correct, or is the actual boundary somewhere different?
Not looking for a ruling on a specific method yet, just trying not to over-read the SentencePiece/pre-tokenized-shard examples before going further.