MiDashengLM-Gen RL-Round2-v2 (rl-r2v2)
Single clean Rank-64 Flow-DPO + Stop-Head RL LoRA anchored directly to ModelsLab/midashenglm-gen-wer-lora (scene-v8).
Paired 132-Clip A/B vs scene-v8
- Blind Gemini 3 Flash overall win-rate: 75.0% (36β12), 0 real regressions
- Ambience: 87.5% (7β1) blind win-rate
- Music: 80.0% (4β1) blind win-rate, CLAP
0.481 -> 0.522(+0.041) - Multi-Tag Mixed: 78.3% (18β5) blind win-rate, WER
26.7% -> 22.3% - SFX: 58.3% (7β5) blind win-rate, score
0.660 -> 0.682 - English Speech WER:
8.99% -> 2.74%
Load stacked on top of scene-v8: ADAPTER_DIR="adapters/scene-v8,adapters/rl-r2v2".
Inference Providers NEW
This model isn't deployed by any Inference Provider. π Ask for provider support
Model tree for ModelsLab/midashenglm-gen-rl-r2v2
Base model
mispeech/midashenglm-gen