Commonsense Reasoning on HellaSwag (Acc %)
70.16HellaSwag AccuracyBefore RL
Evaluation results
| Method | Links | |
|---|---|---|
| Before RLModel=Qwen3-30B-A3B2026.05 | 70.16 | |
| Vanilla RLModel=Qwen3-30B-A3B2026.05 | 69.53 | |
| RL+SAEModel=Qwen3-30B-A3B2026.05 | 69.21 | |
| RL+SAEModel=Qwen3-8B2026.05 | 61.97 | |
| Before RLModel=Qwen3-8B2026.05 | 61.02 | |
| Vanilla RLModel=Qwen3-8B2026.05 | 60.85 | |
| RL+SAEModel=Qwen3-1.7B2026.05 | 40.93 | |
| Vanilla RLModel=Qwen3-1.7B2026.05 | 39.99 | |
| Before RLModel=Qwen3-1.7B2026.05 | 39.66 |