Agent Action Prediction on Switchboard 1s bidirectional event conditions (test)
98.5Weighted F1 (S/H)DualTurn (A - LoRA)
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| DualTurn (A - LoRA)Fine-tuning strategy=LoRA, Event condition duration=1s2026.03 | 98.5 | 97.9 | 76.4 | 86.4 | |
| DualTurn (E - Full FT)Fine-tuning strategy=Full FT, Event condition duration=1s2026.03 | 98.2 | 98.5 | 77.1 | 86.9 | |
| VAP (our eval)Evaluation protocol=fixed-split re-evaluation, Event condition duration=1s2026.03 | 84.3 | 91.6 | 72 | 83.8 |