Science Question Answering on ARC Challenge (ET, ES, E*S, AR%)
45E*SRecover-LoRA
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| Recover-LoRAModel=Qwen3-4B, Quantization=W2-GateUp, Training Data Source=Synthetic-10k, Training Samples=10k2026.06 | 45 | — | — | 82.6 | |
| Recover-LoRAModel=Qwen3-4B, Quantization=W2-GateUp, Training Data Source=OpenHermes-10k, Training Samples=10k2026.06 | 43.2 | — | — | 76.9 | |
| Quantized Student (W2-GateUp)Model=Qwen3-4B, Quantization=W2-GateUp (gate/up at INT2)2026.06 | — | — | 19 | — | |
| Teacher (BF16)Model=Qwen3-4B, Precision=BF162026.06 | — | 50.4 | — | — |