Corrective Instruction Generation on HumanML3D (test)
100BLEUGround-Truth
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| Ground-Truth2024.12 | 100 | 100 | 100 | 1 | |
| CigTimeMode=Full-parameter tuning2024.12 | 24 | 35 | 52 | 0.82 | |
| Mistral-7BMode=In-context learning2024.12 | 16 | 30 | 46 | 0.8 | |
| Llama-3-8BMode=In-context learning2024.12 | 15 | 29 | 45 | 0.77 | |
| Llama-3-8B-LoRAMode=LoRA fine-tuning2024.12 | 10 | 19 | 36 | 0.77 | |
| Mistral-7B-LoRAMode=LoRA fine-tuning2024.12 | 8 | 19 | 27 | 0.79 | |
| MotionGPTInput Template=Section 3.32024.12 | 2 | 10 | 11 | 0.76 | |
| MotionGPT-M2TDescription=Motion captioning baseline2024.12 | 2 | 13 | 12 | 0.76 |