Judgmental event forecasting on Polymarket (val)
0.305NCE LossQwen 3B
Evaluation Results
| Method | Links | |
|---|---|---|
| Qwen 3BTraining stage=End, Post-training method=GRPO2026.06 | 0.305 | |
| Qwen 7BTraining stage=End, Post-training method=GRPO2026.06 | 0.406 | |
| Qwen 1.5BTraining stage=End, Post-training method=GRPO2026.06 | 0.415 | |
| Claude 3.5 SonnetReference model comparison=true2026.06 | 0.62 | |
| Qwen 3BTraining stage=Start, Post-training method=GRPO2026.06 | 0.975 | |
| Qwen 7BTraining stage=Start, Post-training method=GRPO2026.06 | 0.995 | |
| Qwen 1.5BTraining stage=Start, Post-training method=GRPO2026.06 | 1.661 |