Skip to content

Score the merged Whisper checkpoint on the FLEURS monolingual controls #2

Description

@Awshesh12

Table 5 in the paper has three empty cells. Awshesh12/whisper-large-v3-ja-en-cs-merged has never been scored on the monolingual control sets, so its forgetting profile is unknown.

This matters more than a cosmetic gap: the paper argues the LFM adaptation preserves Japanese (CER 6.6 → 6.4) while Whisper's profile is simply unmeasured. A reviewer will ask whether the synthetic fine-tune damaged Whisper's Japanese or English, and we cannot answer.

What's needed

Run 04-evaluation/eval-ft/eval/run_whisper_hf.py over both control sets:

python eval/run_whisper_hf.py --model-id Awshesh12/whisper-large-v3-ja-en-cs-merged \
  --manifest data/eval/fleurs_ja_jp_test200.jsonl --out artifacts/preds/ftwhisper_ja.jsonl
python eval/run_whisper_hf.py --model-id Awshesh12/whisper-large-v3-ja-en-cs-merged \
  --manifest data/eval/fleurs_en_us_test200.jsonl --out artifacts/preds/ftwhisper_en.jsonl
python eval/score.py ftwhisper_ja:artifacts/preds/ftwhisper_ja.jsonl

Use the standard protocol (auto-lang, repetition_penalty=1.3, no_repeat_ngram_size=3) so the numbers are comparable to the zero-shot Whisper row.

Done when

JA-CER, JA-WER and EN-WER are filled in for that row in paper/main.tex Table 5, and the footnote is removed.

Metadata

Metadata

Assignees

Labels

evaluationEvaluation harness, metrics, benchmarkspaper-blockerMust be resolved before the paper can be submitted

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions