Problem-solving on GSM8K (exact match %)
78.94Exact Match AccuracyDS2-INSTRUCT
Evaluation Results
| Method | Links | |
|---|---|---|
| DS2-INSTRUCTModel Family=Qwen2.52026.03 | 78.94 | |
| IA2 onlyBackbone=Qwen3-4B-Base, N=42025.09 | 77.4 | |
| ExploreInstructModel Family=Qwen2.52026.03 | 76.67 | |
| ICLBackbone=Qwen3-4B-Base, N=42025.09 | 76.4 | |
| InstructMixModel Family=Qwen2.52026.03 | 75.42 | |
| IA2 → SFTBackbone=Qwen3-4B-Base, N=42025.09 | 73.6 | |
| Self-InstructModel Family=Qwen2.52026.03 | 72.38 | |
| Zero-ShotModel Family=Qwen2.52026.03 | 71.53 | |
| SFT onlyBackbone=Qwen3-4B-Base, N=42025.09 | 70.9 | |
| DS2-INSTRUCTModel Family=Llama32026.03 | 58.34 | |
| w/o ICLBackbone=Qwen3-4B-Base, N=42025.09 | 56.4 | |
| DS2-INSTRUCTModel Family=Mistral2026.03 | 42.75 | |
| InstructMixModel Family=Mistral2026.03 | 38.26 | |
| InstructMixModel Family=Llama32026.03 | 35.64 | |
| ExploreInstructModel Family=Llama32026.03 | 33.81 | |
| Self-InstructModel Family=Llama32026.03 | 19.83 | |
| Self-InstructModel Family=Mistral2026.03 | 18.52 | |
| ExploreInstructModel Family=Mistral2026.03 | 17.93 | |
| Zero-ShotModel Family=Mistral2026.03 | 14.75 | |
| Zero-ShotModel Family=Llama32026.03 | 12.6 |