List your company

Supervised Fine-Tuning on Heterogeneous SFT Benchmarks (Evaluation)

65GSM8K ScoreEPI
52.62455.83759.0562.263Apr 15, 2026

Evaluation results

MethodLinks
EPIBase Model=Gemma-2-9B2026.046533.269.18.99.18.57
EPIBase Model=Qwen2-7B2026.0463.732.667.28.78.98.23
Static IsolationBase Model=Gemma-2-9B, isolation ratio (p)=1%2026.0462.231.566.48.68.88.21
EPIBase Model=LLaMA-3-8B2026.0461.831.265.48.48.67.98
Static IsolationBase Model=Qwen2-7B, isolation ratio (p)=1%2026.046130.864.98.38.67.93
Multi-Stage (Heuristic)Base Model=Gemma-2-9B2026.0460.430.665.28.28.47.97
Multi-Stage (Random)Base Model=Gemma-2-9B2026.0459.829.864.68.58.77.9
EPIBase Model=Mistral-7B2026.0459.329.563.18.28.47.74
Static IsolationBase Model=LLaMA-3-8B, isolation ratio (p)=1%2026.0459.229.4638.18.37.68
Multi-Stage (Heuristic)Base Model=Qwen2-7B2026.0459.13063.77.98.27.71
Multi-Stage (Random)Base Model=Qwen2-7B2026.0458.529.2638.28.47.67
Full SFTBase Model=Gemma-2-9B2026.0458.530.1648.38.57.78
Multi-Stage (Heuristic)Base Model=LLaMA-3-8B2026.0457.428.762.17.787.48
Full SFTBase Model=Qwen2-7B2026.0457.229.662.488.37.55
Multi-Stage (Random)Base Model=LLaMA-3-8B2026.0456.927.961.588.27.44
Static IsolationBase Model=Mistral-7B, isolation ratio (p)=1%2026.0456.827.961.47.88.17.44
Full SFTBase Model=LLaMA-3-8B2026.0455.628.360.87.88.17.32
Multi-Stage (Heuristic)Base Model=Mistral-7B2026.045527.160.27.47.87.21
Multi-Stage (Random)Base Model=Mistral-7B2026.0454.426.359.67.787.17
Full SFTBase Model=Mistral-7B2026.0453.126.758.97.57.97.05