Instruction Tuning on Vicuna
18.73RougeL ScoreARMADA
Evaluation Results
| Method | Links | |
|---|---|---|
| ARMADATeacher=text-to-image, Loss Function=L_cosine, Student Model=LLaMA-3.2-3B2026.03 | 18.73 | |
| ARMADATeacher=text-to-audio, Loss Function=L_elementwise, Student Model=LLaMA-3.2-3B2026.03 | 18.67 | |
| ARMADATeacher=text-to-video, Loss Function=L_cosine, Student Model=LLaMA-3.2-3B2026.03 | 18.44 | |
| ARMADATeacher=text-to-image, Loss Function=L_elementwise, Student Model=LLaMA-3.2-3B2026.03 | 18.16 | |
| ARMADATeacher=text-to-image, Loss Function=L_euclid, Student Model=LLaMA-3.2-3B2026.03 | 18.15 | |
| ARMADATeacher=text-to-audio, Loss Function=L_cosine, Student Model=LLaMA-3.2-3B2026.03 | 18.15 | |
| ARMADATeacher=text-to-audio, Loss Function=L_euclid, Student Model=LLaMA-3.2-3B2026.03 | 18.01 | |
| ARMADATeacher=text-to-video, Loss Function=L_euclid, Student Model=LLaMA-3.2-3B2026.03 | 17.92 | |
| ARMADATeacher=text-to-video, Loss Function=L_elementwise, Student Model=LLaMA-3.2-3B2026.03 | 17.84 | |
| SeqKDTeacher=language-only, Distillation Type=SeqKD, Student Model=LLaMA-3.2-3B2026.03 | 17.74 | |
| UndistilledTeacher=-, Distillation Type=undistilled, Student Model=LLaMA-3.2-3B2026.03 | 17.59 | |
| S2FTModel=LLaMA2-13B, # Trainable Parameters=0.08M2026.05 | 7.94 | |
| FourierFTModel=LLaMA2-13B, # Trainable Parameters=0.08M2026.05 | 7.92 | |
| LoRAModel=LLaMA2-13B, # Trainable Parameters=250.3M, Updating layers other than lm_head=true2026.05 | 7.89 | |
| LoRAModel=LLaMA2-13B, # Trainable Parameters=52.4M, Updating layers other than lm_head=false2026.05 | 7.89 | |
| S2FTModel=LLaMA2-7B, # Trainable Parameters=0.064M2026.05 | 7.5 | |
| FourierFTModel=LLaMA2-7B, # Trainable Parameters=0.064M2026.05 | 7.49 | |
| LoRAModel=LLaMA2-7B, # Trainable Parameters=159.9M, Updating layers other than lm_head=true2026.05 | 7.38 | |
| LoRAModel=LLaMA2-7B, # Trainable Parameters=33.5M, Updating layers other than lm_head=false2026.05 | 7.35 |