Skip to content

fix(distill): byt5 decode_joined mojibake — all Arabic labels poisoned - #35

Merged
ronaldtse merged 6 commits into
mainfrom
fix/byt5-decode
Aug 24, 2026
Merged

fix(distill): byt5 decode_joined mojibake — all Arabic labels poisoned#35
ronaldtse merged 6 commits into
mainfrom
fix/byt5-decode

Conversation

@ronaldtse

Copy link
Copy Markdown
Contributor

Summary

  • byt5 under transformers 5.14: convert_ids_to_tokens returns byte tokens as raw chars, so decode_joined double-encodes — every Arabic teacher label was mojibake; both Arabic students trained on corrupted targets (their identical 83.08 DERs were the bare-text constant)
  • decode_joined now branches to batch_decode for byte-level vocabs (byte-exact, asserted); umt5 path unchanged
  • ara-diac-small relabels onto teacher_labels_v2.jsonl; the ara-tiny capacity verdict is requalified as confounded (Thai conclusion unaffected)

Test plan

  • CI green
  • relabeled ara-diac-small trains and decodes real Arabic

Ronald Tse added 3 commits August 24, 2026 06:28
…decode

convert_ids_to_tokens on a byte-level vocab returns each byte token as
a raw char; joining them double-encodes the text. This poisoned every
Arabic label generated under 5.14 (both students trained on mojibake
targets — their identical 83.08 DER was the bare-text constant, not a
capacity verdict). Byte-exactness asserted locally; umt5 path
unchanged. ara-diac-small relabels onto teacher_labels_v2.jsonl.
Comment thread src/api/inference.py Fixed
@ronaldtse
ronaldtse merged commit 683337a into main Aug 24, 2026
10 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants