docs: one source for the numbers, and fix which whisper build is default - #3
Merged
Merged
Conversation
…r build is default Every model README carried its numbers twice: once in a table somebody typed, once in the generated `facts` block below it. The generated one is rewritten from the registry; the typed one is not, and it drifted exactly as the script's own docstring predicted — after the registry moved gipformer's Vietnamese WER to 8.5 %, the prose above still said 2.41 % in both languages, in the same file. The typed tables are removed. Facts about the artefact come from the manifest, and the prose keeps what no manifest can state: why the model is in the product. `whisper-tiny`'s README also called int8 "the default variant". It is not, and the measurement says it should not be: on 100 FLEURS clips the two builds are the same speed and int8 is 13.7 WER points worse, so `variant::choose` prefers fp32 and drops to int8 only when memory forces it. The README now says that, with the numbers and the reason the encoder-vs-decoder split produces them.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The typed tables had already drifted
Every model README stated its metrics twice — a hand-written table, and the generated
factsblockbelow it. Only the second is rewritten from the registry. After
summo-registry#2moved gipformer'sVietnamese WER to 8.5 %, the same file said both:
The typed tables are deleted from all eight READMEs.
scripts/readme.py --checkpasses.int8 was labelled
whisper-tiny's default variantIt is not, and it should not be. 100 FLEURS vi clips, same runtime and machine:
Same speed, 13.7 points worse.
variant::choosealready ranks fp32 first and falls back to int8only when memory forces it; the README now agrees with the code, and explains why — on the encoder
alone int8 is 1.24× faster, and the autoregressive decoder gives it all back.
🤖 Generated with Claude Code