Audio Generation on EpicBench T2A 1.0 (test)
68Win Rate EOST2A-Feedback + DPO
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| T2A-Feedback + DPOScoring Protocol=Human Scoring, Tuning Strategy=Direct Preference Optimization (DPO), Tuning Dataset=T2A-Feedback, Base Model=Make-an-Audio 22025.05 | 68 | 62 | 68 | |
| Audio-Alpaca + DPOScoring Protocol=Human Scoring, Tuning Strategy=Direct Preference Optimization (DPO), Tuning Dataset=Audio-Alpaca, Base Model=Make-an-Audio 22025.05 | 65 | 64 | 59 | |
| T2A-Feedback + RAFTScoring Protocol=Human Scoring, Tuning Strategy=Reward rAnked FineTuning (RAFT), Tuning Dataset=T2A-Feedback, Base Model=Make-an-Audio 22025.05 | 61 | 57 | 61 | |
| T2A-Feedback + DPOScoring Protocol=AI Scoring, Tuning Strategy=Direct Preference Optimization (DPO), Tuning Dataset=T2A-Feedback, Base Model=Make-an-Audio 22025.05 | 58 | 64 | 52 | |
| Audio-Alpaca + RAFTScoring Protocol=Human Scoring, Tuning Strategy=Reward rAnked FineTuning (RAFT), Tuning Dataset=Audio-Alpaca, Base Model=Make-an-Audio 22025.05 | 57 | 54 | 53 | |
| Audio-Alpaca + DPOScoring Protocol=AI Scoring, Tuning Strategy=Direct Preference Optimization (DPO), Tuning Dataset=Audio-Alpaca, Base Model=Make-an-Audio 22025.05 | 55 | 52 | 49 | |
| Audio-Alpaca + RAFTScoring Protocol=AI Scoring, Tuning Strategy=Reward rAnked FineTuning (RAFT), Tuning Dataset=Audio-Alpaca, Base Model=Make-an-Audio 22025.05 | 53 | 51 | 42 | |
| T2A-Feedback + RAFTScoring Protocol=AI Scoring, Tuning Strategy=Reward rAnked FineTuning (RAFT), Tuning Dataset=T2A-Feedback, Base Model=Make-an-Audio 22025.05 | 52 | 52 | 54 |