hero3
Step 1
| Rollout | Reward | Test | Edited accuracy | Hit cap | Tensors edited | Turns | Minutes | Forward passes | Python calls | Ended |
|---|
| 1 | +1.2 | +0.6 | 8.6% | 92.6% | 1 | 76 | 105.6 | 15 | 55 | editor finished |
| 2 | +9.4 | +11.1 | 16.8% | 82.4% | 48 | 66 | 52.7 | 7 | 54 | editor finished |
| 3 | +18.8 | +16.7 | 26.2% | 78.5% | 6 | 87 | 88.4 | 11 | 73 | editor finished |
| 4 | +15.6 | +15.2 | 23.0% | 75.4% | 6 | 77 | 85.9 | 11 | 63 | editor finished |
| 5 | +33.6 | +32.8 | 41.0% | 39.8% | 48 | 84 | 66.1 | 9 | 72 | editor finished |
| 6 | +0.0 | no edit | โ | โ | 0 | 27 | 8.6 | 0 | 19 | no tool call in five replies |
| 7 | +2.7 | +2.5 | 10.2% | 90.6% | 1 | 96 | 106.2 | 14 | 79 | editor finished |
| 8 | +27.3 | +29.4 | 34.8% | 58.6% | 48 | 86 | 59.9 | 14 | 69 | editor finished |
| Mean | +13.6 | +15.5 | 22.9% | 74.0% | 19.8 | 74.9 | 71.7 | 10.1 | 60.5 | |
Step 2
| Rollout | Reward | Test | Edited accuracy | Hit cap | Tensors edited | Turns | Minutes | Forward passes | Python calls | Ended |
|---|
| 1 | +32.0 | +31.4 | 39.5% | 55.1% | 48 | 79 | 62.9 | 13 | 62 | editor finished |
| 2 | +20.7 | +22.0 | 28.1% | 64.5% | 48 | 67 | 82.6 | 14 | 51 | editor finished |
| 3 | +13.3 | +13.4 | 20.7% | 77.7% | 48 | 42 | 58.0 | 7 | 32 | editor finished |
| 4 | +25.8 | +24.2 | 33.2% | 59.0% | 48 | 34 | 49.5 | 5 | 27 | editor finished |
| 5 | +24.2 | +22.4 | 31.6% | 65.6% | 48 | 75 | 78.5 | 12 | 60 | editor finished |
| 6 | +27.3 | +28.3 | 34.8% | 53.9% | 48 | 64 | 110.8 | 13 | 48 | editor finished |
| 7 | +23.1 | +24.2 | 30.5% | 60.2% | 48 | 52 | 49.4 | 9 | 40 | editor finished |
| 8 | +27.7 | +28.1 | 35.2% | 53.1% | 48 | 42 | 49.6 | 7 | 32 | editor finished |
| Mean | +24.3 | +24.2 | 31.7% | 61.1% | 48.0 | 56.9 | 67.7 | 10.0 | 44.0 | |
Reward: edited minus original accuracy on the reward set (256 answers, original 7.4%), in points. Test: the same on MATH-500 (original 6.6%), not used for training; scored after a step finishes, โ until then. Hit cap: share of answers reaching 8,192 tokens (original 93%).