ci: bench remote (officiels) — transcription API + WER#37
Merged
Conversation
Workflow workflow_dispatch : pull corpus public HF, transcription des modeles remote (Albert, Voxtral/Scaleway, ElevenLabs en option), scoring WER, push des resultats vers AgentPublic/eval-stt-results. Modeles locaux exclus (phase 2, runner self-hosted). Skips Voxtral temporaires (> 28 min + silence connu). Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
|
Note Gemini is unable to generate a review for this pull request due to the file types involved not being currently supported. |
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Objectif
Câbler en CI la transcription des discours officiels (corpus public
AgentPublic/eval-stt-officiels) par les modèles accessibles via API, puis le calcul du WER. Première brique du chantier CI/CD des évaluations Transcript.Déclenchement manuel (
workflow_dispatch) : chaque run consomme des appels API facturables.Ce que fait le workflow
score all, normalisationstandard_numbers), rapport horodaté + révision du corpus estampillée.AgentPublic/eval-stt-results, et artefacts GitHub.Les modèles locaux (WhisperX, Kyutai, Cohere/MLX) ne sont pas ici : ils exigent un GPU ou du MLX et passeront par un runner self-hosted (phase 2).
Périmètre Voxtral (limitations temporaires)
Voxtral a une limite de 30 min/fichier et collapse sur certains silences longs :
ffprobe) ; un long sample ramené sous la limite est couvert sans toucher la CI ;crise_energetique_aides_mai(silence long connu).Exclusions temporaires : le vrai fix est de porter le chunk VAD dans le provider
scaleway(à faire dans la reprise de #36).Prérequis avant que ça tourne
Secrets à poser dans Settings > Secrets and variables > Actions :
ALBERT_API_KEYSCW_SECRET_KEY,SCW_DEFAULT_PROJECT_IDELEVENLABS_API_KEY(si on inclut ElevenLabs)HF_TOKEN(write sur AgentPublic, pour pousser les résultats ; sans lui le push est sauté et seuls les artefacts sont produits)Tant que les secrets ne sont pas posés, le job échoue : PR laissée en draft.
Notes de conception
🤖 Generated with Claude Code