RL — RAFT on Recipe I (Reward rAnked FineTuning). Tag F1 28%, Tag Recall 50%, -5pp hallucination vs SFT. Pair with base for ASR; see serve_modal.py Mode B hybrid. https://github.com/YongkangZOU/evoxtral-realtime

acfd33d

verified

YongkangZOU commited on May 2