Qwen3.8-27B Nemotron VI+EN Mixed ASR + QA
AudioLLM checkpoint trained from the dedicated Nemotron Vietnamese+English E2E ASR checkpoint with a mixed training recipe containing:
- ASR transcription data
- QA records with system instructions
- QA records without system instructions
The Nemotron encoder is frozen and the 1024 -> 5120 latent projector is trained. The base Qwen3.8-27B model is downloaded separately by the local Luvox AudioLLM loader.
Training config: training_recipe.yaml
Model config: model_config.yaml
Audio rate: 12.5 Hz Nemotron features
Training limits: 0.2 s minimum audio, 20 s maximum audio, ASR 1024 text tokens, QA 768 text tokens
Validation: ASR WER/CER on LSVSC, PW733, and GigaSpeech
The best repository is the checkpoint selected by validation CER. The latest repository is the most recent training state converted to a deployable model-only checkpoint.
Checkpoint: latest