Warisan2026/Yoruba-ATES-Evaluation

Fix transcript alignment API to accept mfa_isolation_dir

Add fusion 7-context SSL predictor and deploy checkpoint

Revert "Add batch BiLSTM JSON ZIP support"

Revert "Simplify: remove BiLSTM JSON upload; default batch max_utterances=200"

Revert "Fix BiLSTM Yoruba vowel symbols; improve quality score breakdown"

Revert "Fix voiced_frac calculation; improve no-segment error diagnostics"

Revert "Fix MFA OOV from <unk> placeholders"

Revert "Make batch compare async to avoid timeouts"

Revert "Handle non-JSON responses and batch endpoint fallback"

Revert "Fix IRR JSON handling and log download"

Revert "Return JSON for API 500s and improve batch status errors"

Revert "Fix batch job status 500 and cache races"

Revert "Avoid extracting entire zips in batch compare"

Revert "Add downloadable batch log file and cleanup"

Revert "Capture batch logs client-side and allow download"

Revert "Make batch MFA single-run + persist batch logs/state"

Revert "Add batch live log viewer and reduce batch memory"

Revert "Revert "Add batch live log viewer and reduce batch memory""

Revert "Add batch live log viewer and reduce batch memory"

Add batch live log viewer and reduce batch memory

Make batch MFA single-run + persist batch logs/state

Capture batch logs client-side and allow download

Add downloadable batch log file and cleanup

Avoid extracting entire zips in batch compare

Fix batch job status 500 and cache races

Return JSON for API 500s and improve batch status errors

Fix IRR JSON handling and log download

Handle non-JSON responses and batch endpoint fallback

Make batch compare async to avoid timeouts

Fix MFA OOV from <unk> placeholders

Fix voiced_frac calculation; improve no-segment error diagnostics

Fix BiLSTM Yoruba vowel symbols; improve quality score breakdown

Simplify: remove BiLSTM JSON upload; default batch max_utterances=200

Add batch BiLSTM JSON ZIP support

Fix batch compare to honor tone_predictor; improve BiLSTM diagnostics

Update technical approach updates with implemented statistical rigor

Add statistical rigor: batch significance, IRR endpoint, interpretation guidance

Add technical approach updates summary

Add actual audio files for samples 1000, 1001, 1002

Fix sample transcription mismatch and add new audio samples

Fix MFA filename mismatch issue - auto-rename TextGrids when MFA uses wrong filename

Fix sample transcript to match actual audio content - update to correct Yoruba text

Fix sample data loading - ensure workflow switching and correct file paths

Add sample data selector for quick testing

Add OOV handler for automatic lexicon extension

Return downloadable logs when MFA produces no TextGrid

Install newer sqlite to support UPDATE FROM

Make apt installs noninteractive (tzdata)

Install libxml2 so mamba works on HF builder

Trigger HF rebuild