bosonai/higgs-audio-v3-8b-stt

Remove v12_pf_ab01_7dom_dro/ — recipe failed at full scale (AMI 11.00, E22 10.76; killed by allocator). 5.33% 500-sample claim was severely optimistic.

Revert v12 README to original (per-user request)

Correct v12 README: 5.33% is 500-sample probe; project ~6.0% full-scale (NOT champion-beater)

Add open_asr_leaderboard/v12_pf_ab01_7dom_dro/ — 7-domain Group DRO LoRA (internal 5.33%, scorer caveat per PR#135)

Add open_asr_leaderboard/v12_pf_ab01_7dom_dro/ — 7-domain Group DRO LoRA (internal 5.33%, scorer caveat per PR#135)

Add open_asr_leaderboard/v12_pf_ab01_7dom_dro/ — 7-domain Group DRO LoRA (internal 5.33%, scorer caveat per PR#135)

Upload v11 AMI 10x LoRA adapter (AMI 9.54%, LS-clean 1.27% — did not beat baseline 9.36%)

Add v10 LoRA: adapter_model.safetensors

Add v10 LoRA: adapter_config.json

Add leaderboard reproduction helpers

Add leaderboard reproduction helpers

Add leaderboard reproduction helpers

Add open-asr-leaderboard reproduction script (5.30% 8B / 5.67% 1.7B)

Embed internal leaderboard-parity eval logic in transcribe.py

fix: also suppress right-padding warning from transformers logger

fix: revert pad_left to False (model expects right-padded input + re-pads internally)

fix: use pad_left=True in collator (decoder-only batched generation needs left-padding)

feat: batched generation for ASR (remove bs=1 assertion when no audio_out tokens)

feat: batched generation for ASR (remove bs=1 assertion when no audio_out tokens)

feat: add transcribe_batch for parallel inference (single transcribe wraps it)

fix: extract transcription from open <think> block (model does not emit </think>)

fix: use **batch for generate, handle tuple output, proper think-block parsing

fix: add all missing attrs to ChatMLDatasetSample (reward, group_id, get_audio_codes, etc)

fix: add missing audio_label_ids_concat + num_audios to ChatMLDatasetSample

fix: dual-mode imports + add get_wv/get_audio_cuts to ChatMLDatasetSample

fix: dual-mode imports + add get_wv/get_audio_cuts to ChatMLDatasetSample

fix: dual-mode imports + add get_wv/get_audio_cuts to ChatMLDatasetSample

fix: dual-mode imports + add get_wv/get_audio_cuts to ChatMLDatasetSample

fix: add auto_map + make xcodec imports lazy for trust_remote_code support

fix: add auto_map + make xcodec imports lazy for trust_remote_code support

fix: add auto_map + make xcodec imports lazy for trust_remote_code support

Update usage: standalone inference without boson_multimodal dependency

Add transcribe.py — standalone inference without boson_multimodal

Add higgs_audio_collator.py — standalone inference without boson_multimodal

Clean up model card — remove training details

Update README: 5.30% avg WER (new best), updated per-dataset results and training info

Upload merged model weights (LoRA idea-0524ba merged): 5.30% avg WER

Upload LoRA weights (idea-0524ba): 5.30% avg WER, rank 64, MLP targeting

Update ESB benchmark results: 5.30% avg WER (new best, beats 5.40% target)

Add ESB benchmark eval results (model-index) for Open ASR Leaderboard

Remove auto_map to fix vLLM 0.10.2 compatibility

Add README with model card and benchmark results

Upload Higgs Audio v3 8B STT (merged LoRA v7)

initial commit