Language Understanding on MMLU-ProX T5
47.1AccuracyFull Retraining
Evaluation results
| Method | Links | |
|---|---|---|
| Full RetrainingBackbone=Phi-4-Mini-Instruct-3.8B2026.04 | 47.1 | |
| COMPASS-ECDABackbone=Phi-4-Mini-Instruct-3.8B2026.04 | 47 | |
| EWCBackbone=Phi-4-Mini-Instruct-3.8B2026.04 | 44 | |
| Random RehearsalBackbone=Phi-4-Mini-Instruct-3.8B2026.04 | 43.8 | |
| COMPASS-ECDAModel=Qwen2.5-7B-Instruct2026.04 | 43.7 | |
| Full RetrainingModel=Qwen2.5-7B-Instruct2026.04 | 42.7 | |
| Naive Fine-tuningModel=Qwen2.5-7B-Instruct2026.04 | 42 | |
| EWCModel=Qwen2.5-7B-Instruct2026.04 | 41.8 | |
| Random RehearsalModel=Qwen2.5-7B-Instruct2026.04 | 41.7 | |
| Naive Fine-tuningBackbone=Phi-4-Mini-Instruct-3.8B2026.04 | 31.9 | |
| COMPASS-ECDABackbone=Phi-4-Mini-Instruct-3.8B2026.04 | 28.1 | |
| Full RetrainingBackbone=Phi-4-Mini-Instruct-3.8B2026.04 | 27.6 | |
| Random RehearsalBackbone=Phi-4-Mini-Instruct-3.8B2026.04 | 27.4 | |
| Naive Fine-tuningBackbone=Phi-4-Mini-Instruct-3.8B2026.04 | 27 | |
| EWCBackbone=Phi-4-Mini-Instruct-3.8B2026.04 | 26.5 |