Task Gradient Conflict on Sequential Fine-Tuning Suite (CodeAlpaca, LogiQA, Alpaca) (test)
13Code Conflict ScoreEPI
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| EPIBase Model=Gemma-2-9B, p=1%2026.04 | 13 | 11 | 9 | 8 | 7 | 9.6 | |
| Static IsolationBase Model=Gemma-2-9B, p=1%2026.04 | 14 | 12 | 11 | 9 | 8 | 10.8 | |
| EPIBase Model=LLaMA-3-8B, p=1%2026.04 | 15 | 12 | 10 | 9 | 8 | 10.8 | |
| EPIBase Model=Qwen2-7B, p=1%2026.04 | 16 | 13 | 11 | 9 | 8 | 11.4 | |
| Static IsolationBase Model=Qwen2-7B, p=1%2026.04 | 17 | 15 | 13 | 11 | 10 | 13.2 | |
| Static IsolationBase Model=LLaMA-3-8B, p=1%2026.04 | 18 | 15 | 13 | 12 | 11 | 13.8 | |
| Multi-Stage (Heuristic)Base Model=Gemma-2-9B2026.04 | 25 | 23 | 21 | 19 | 18 | 21.2 | |
| Multi-Stage (Heuristic)Base Model=LLaMA-3-8B2026.04 | 28 | 25 | 23 | 21 | 20 | 23.4 | |
| Multi-Stage (Random)Base Model=Gemma-2-9B2026.04 | 28 | 26 | 23 | 22 | 20 | 23.8 | |
| Multi-Stage (Heuristic)Base Model=Qwen2-7B2026.04 | 30 | 27 | 25 | 23 | 20 | 25 | |
| Multi-Stage (Random)Base Model=LLaMA-3-8B2026.04 | 31 | 28 | 26 | 24 | 22 | 26.2 | |
| Multi-Stage (Random)Base Model=Qwen2-7B2026.04 | 33 | 30 | 27 | 25 | 22 | 27.4 | |
| Full SFT (Multi-task)Base Model=Gemma-2-9B2026.04 | 39 | 36 | 33 | 30 | 28 | 33.2 | |
| Full SFT (Multi-task)Base Model=LLaMA-3-8B2026.04 | 42 | 38 | 36 | 33 | 30 | 35.8 | |
| Full SFT (Multi-task)Base Model=Qwen2-7B2026.04 | 44 | 40 | 37 | 35 | 32 | 37.6 |