Open-ended Instruction Following on AlpacaEval Audio
4.59EN ScoreSEA-LION-v3-8B-IT
Evaluation results
| Method | Links | ||
|---|---|---|---|
| SEA-LION-v3-8B-ITApproach category=Text-only reference2026.03 | 4.59 | 4.21 | |
| MERaLiON-2-10BApproach category=Zero-shot end-to-end2026.03 | 4.59 | 1.96 | |
| Ours (soft-gating)Approach category=Distilled end-to-end, Gating strategy=soft-gating2026.03 | 4.58 | 3.21 | |
| Ours (hard-gating)Approach category=Distilled end-to-end, Gating strategy=hard-gating2026.03 | 4.39 | 3.33 | |
| ML-DiVAApproach category=Distilled end-to-end, Training protocol=re-trained baseline2026.03 | 4.28 | 2.87 | |
| Whisper-v3 + SEA-LION-v3-8B-ITApproach category=Cascaded2026.03 | 3.99 | 3.95 | |
| SeaLLMs-AudioApproach category=Zero-shot end-to-end2026.03 | 3.93 | 3.72 | |
| Qwen2-AudioApproach category=Zero-shot end-to-end2026.03 | 3.77 | 3.65 | |
| Glm-4-voiceApproach category=Zero-shot end-to-end2026.03 | 3.62 | 3.41 | |
| EN-DiVAApproach category=Distilled end-to-end, Training protocol=re-trained baseline2026.03 | 3.57 | — |