EleutherAI

/

pythia-2.8b

Like 35

Follow

EleutherAI 1.39k

While measuring across this repository's revision labels, we found two artifact-level issues by checking file digests across the repository's labels:

1. Nine revision labels — step54000, step56000 through step59000, and step61000 through step64000 — are byte-identical to step143000 in both the monolithic model.safetensors and pytorch_model.bin files. step55000 and step60000 are not members of this duplicate group. The digest evidence establishes the aliasing but not which training step the shared bytes actually represent.
2. A different monolithic model.safetensors digest appears at 77 revision labels. The sharded safetensors (when present) or the PyTorch binary avoids that specific stale monolithic-file digest. However, the same census shows that step26000, step53000, and step103000 resolve to the main commit, so we cannot claim that an alternative file under every one of those 77 labels is label-correct.

Practical workaround we used: prefer the sharded safetensors index, then the PyTorch binary, and only then the monolithic file; verify both file-digest uniqueness and resolved-commit uniqueness across labels. We read step65000 in place of step64000.

Happy to share the per-label digest table if useful.