List your company
ResearchFormerly WizWand

Task Gradient Conflict on Sequential Fine-Tuning Suite (CodeAlpaca, LogiQA, Alpaca) (test)

13Code Conflict ScoreEPI
11.7620.1328.536.87Apr 15, 2026
Updated 6mo ago

Evaluation results

MethodLinks
EPIBase Model=Gemma-2-9B, p=1%2026.0413119879.6
Static IsolationBase Model=Gemma-2-9B, p=1%2026.041412119810.8
EPIBase Model=LLaMA-3-8B, p=1%2026.041512109810.8
EPIBase Model=Qwen2-7B, p=1%2026.041613119811.4
Static IsolationBase Model=Qwen2-7B, p=1%2026.04171513111013.2
Static IsolationBase Model=LLaMA-3-8B, p=1%2026.04181513121113.8
Multi-Stage (Heuristic)Base Model=Gemma-2-9B2026.04252321191821.2
Multi-Stage (Heuristic)Base Model=LLaMA-3-8B2026.04282523212023.4
Multi-Stage (Random)Base Model=Gemma-2-9B2026.04282623222023.8
Multi-Stage (Heuristic)Base Model=Qwen2-7B2026.04302725232025
Multi-Stage (Random)Base Model=LLaMA-3-8B2026.04312826242226.2
Multi-Stage (Random)Base Model=Qwen2-7B2026.04333027252227.4
Full SFT (Multi-task)Base Model=Gemma-2-9B2026.04393633302833.2
Full SFT (Multi-task)Base Model=LLaMA-3-8B2026.04423836333035.8
Full SFT (Multi-task)Base Model=Qwen2-7B2026.04444037353237.6