Continual Instruction Tuning on Five continual instruction tuning sequences
66.9Average Accuracy (AA)PROXYMIX
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| PROXYMIXTarget Model=LLaMA-3-8B, Replay Controller=Pythia-410M2026.05 | 66.9 | 9.3 | 11.4 | |
| PROXYMIXTarget Model=Mistral-7B, Replay Controller=Pythia-410M2026.05 | 65.1 | 8.3 | 10.7 | |
| PROXYMIXTarget Model=Pythia-6.9B, Replay Controller=Pythia-410M2026.05 | 63.4 | 7.8 | 8.1 | |
| Random ReplayTarget Model=LLaMA-3-8B, Replay Controller=Pythia-410M2026.05 | 57.6 | — | — | |
| Random ReplayTarget Model=Mistral-7B, Replay Controller=Pythia-410M2026.05 | 56.8 | — | — | |
| Random ReplayTarget Model=Pythia-6.9B, Replay Controller=Pythia-410M2026.05 | 55.6 | — | — |