rebase2
Step 1 (shared with hero3)
| 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 | – | – | 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 (shared with nolib1)
| Rollout | Reward | Test | Edited accuracy | Hit cap | Tensors edited | Turns | Minutes | Forward passes | Python calls | Ended |
|---|
| 1 | +21.1 | +19.0 | 28.5% | 70.3% | 48 | 81 | 47.1 | 9 | 68 | editor finished |
| 2 | +11.3 | +12.3 | 18.8% | 80.5% | 4 | 69 | 73.0 | 12 | 54 | editor finished |
| 3 | +16.8 | +18.7 | 24.2% | 74.6% | 48 | 61 | 58.8 | 12 | 47 | editor finished |
| 4 | +21.1 | +21.6 | 28.5% | 69.1% | 1 | 91 | 88.8 | 13 | 73 | editor finished |
| 5 | +0.4 | +0.4 | 7.8% | 93.0% | 1 | 63 | 80.2 | 9 | 50 | editor finished |
| 6 | +13.3 | +10.0 | 20.7% | 83.2% | 18 | 72 | 119.5 | 13 | 53 | editor finished |
| 7 | +21.1 | +21.4 | 28.5% | 69.5% | 1 | 84 | 78.0 | 14 | 67 | editor finished |
| 8 | +3.1 | +4.4 | 10.5% | 90.2% | 48 | 68 | 59.0 | 8 | 56 | editor finished |
| Mean | +13.5 | +13.5 | 20.9% | 78.8% | 21.1 | 73.6 | 75.5 | 11.2 | 58.5 | |
Step 3 (shared with nolib1)
| Rollout | Reward | Test | Edited accuracy | Hit cap | Tensors edited | Turns | Minutes | Forward passes | Python calls | Ended |
|---|
| 1 | +32.8 | – | 40.2% | 47.7% | 48 | 86 | 66.6 | 13 | 69 | editor finished |
| 2 | +27.7 | – | 35.2% | 60.2% | 48 | 97 | 85.6 | 15 | 81 | editor finished |
| 3 | +0.0 | – | – | – | 0 | 46 | 13.4 | 4 | 30 | no tool call in five replies |
| 4 | +16.8 | – | 24.2% | 74.6% | 30 | 73 | 51.3 | 10 | 60 | editor finished |
| 5 | +14.4 | – | 21.9% | 80.5% | 6 | 78 | 68.6 | 10 | 66 | editor finished |
| 6 | +18.4 | – | 25.8% | 69.9% | 24 | 97 | 89.3 | 15 | 78 | editor finished |
| 7 | +25.0 | – | 32.4% | 62.5% | 24 | 105 | 119.1 | 18 | 84 | editor finished |
| 8 | +14.4 | – | 21.9% | 76.2% | 48 | 81 | 79.8 | 13 | 65 | editor finished |
| Mean | +18.7 | – | 28.8% | 67.4% | 28.5 | 82.9 | 71.7 | 12.2 | 66.6 | |
Step 4
| Rollout | Reward | Test | Edited accuracy | Hit cap | Tensors edited | Turns | Minutes | Forward passes | Python calls | Ended |
|---|
| 1 | +30.5 | +32.4 | 37.9% | 37.9% | 48 | 68 | 85.2 | 12 | 53 | editor finished |
| 2 | +13.7 | +12.2 | 21.1% | 82.0% | 9 | 52 | 92.8 | 10 | 41 | editor finished |
| 3 | +12.9 | +13.9 | 20.3% | 77.7% | 2 | 120 | 90.7 | 15 | 101 | context limit |
| 4 | +28.5 | +27.5 | 35.9% | 57.4% | 24 | 65 | 72.8 | 10 | 53 | editor finished |
| 5 | +21.5 | +18.6 | 28.9% | 71.1% | 1 | 71 | 115.2 | 11 | 56 | editor finished |
| 6 | +31.2 | +31.2 | 38.7% | 43.4% | 25 | 87 | 124.8 | 15 | 70 | time limit |
| 7 | +23.4 | +23.8 | 30.9% | 65.2% | 11 | 63 | 69.6 | 7 | 55 | editor finished |
| 8 | -0.8 | +0.5 | 6.6% | 94.5% | 3 | 71 | 111.2 | 14 | 54 | editor finished |
| Mean | +20.1 | +20.0 | 27.5% | 66.2% | 15.4 | 74.6 | 95.3 | 11.8 | 60.4 | |
Step 5
| Rollout | Reward | Test | Edited accuracy | Hit cap | Tensors edited | Turns | Minutes | Forward passes | Python calls | Ended |
|---|
| 1 | +18.4 | +15.1 | 25.8% | 74.2% | 24 | 67 | 78.5 | 10 | 53 | editor finished |
| 2 | +4.7 | +4.0 | 12.1% | 88.3% | 1 | 76 | 108.9 | 12 | 62 | editor finished |
| 3 | +29.7 | +26.6 | 37.1% | 59.0% | 24 | 77 | 91.2 | 11 | 64 | editor finished |
| 4 | +24.6 | +27.0 | 32.0% | 66.4% | 24 | 84 | 113.9 | 12 | 68 | editor finished |
| 5 | +10.5 | +13.8 | 18.0% | 80.1% | 1 | 75 | 138.4 | 10 | 62 | time limit |
| 6 | +5.1 | +6.8 | 12.5% | 91.8% | 1 | 57 | 132.6 | 7 | 48 | editor finished |
| 7 | +2.3 | +2.4 | 9.8% | 89.8% | 48 | 72 | 125.6 | 8 | 61 | editor finished |
| 8 | +21.1 | +21.9 | 28.5% | 68.4% | 1 | 80 | 125.7 | 9 | 67 | editor finished |
| Mean | +14.6 | +14.7 | 22.0% | 77.2% | 15.5 | 73.5 | 114.3 | 9.9 | 60.6 | |
Step 6
| Rollout | Reward | Test | Edited accuracy | Hit cap | Tensors edited | Turns | Minutes | Forward passes | Python calls | Ended |
|---|
| 1 | +0.0 | – | – | – | 0 | 118 | 75.2 | 16 | 100 | editor finished |
| 2 | +0.0 | – | – | – | 0 | 119 | 77.4 | 15 | 99 | editor finished |
| 3 | +0.0 | – | – | – | 0 | 82 | 67.4 | 6 | 73 | editor finished |
| 4 | +0.0 | – | 38.3% | 38.7% | 1 | 100 | 69.4 | 10 | 89 | context limit |
| 5 | +0.0 | – | – | – | 0 | 99 | 66.6 | 11 | 84 | editor finished |
| 6 | +0.4 | – | 38.7% | 39.8% | 5 | 107 | 79.3 | 14 | 91 | editor finished |
| 7 | +2.7 | – | 41.0% | 40.2% | 2 | 95 | 60.8 | 10 | 82 | editor finished |
| 8 | +0.0 | – | – | – | 0 | 85 | 52.4 | 9 | 74 | editor finished |
| Mean | +0.4 | – | 39.3% | 39.6% | 1.0 | 100.6 | 68.6 | 11.4 | 86.5 | |
Step 7
| Rollout | Reward | Test | Edited accuracy | Hit cap | Tensors edited | Turns | Minutes | Forward passes | Python calls | Ended |
|---|
| 1 | +0.0 | – | – | – | 0 | 91 | 57.9 | 11 | 77 | editor finished |
| 2 | +0.0 | – | – | – | 0 | 119 | 58.1 | 12 | 103 | editor finished |
| 3 | -1.2 | – | 37.1% | 41.8% | 2 | 132 | 99.2 | 14 | 115 | context limit |
| 4 | +0.0 | – | – | – | 0 | 147 | 51.3 | 16 | 127 | context limit |
| 5 | +5.1 | – | 43.4% | 30.1% | 1 | 81 | 54.3 | 9 | 71 | editor finished |
| 6 | +3.5 | – | 41.8% | 34.0% | 3 | 90 | 57.6 | 10 | 78 | editor finished |
| 7 | +0.0 | – | – | – | 0 | 128 | 52.7 | 10 | 117 | editor finished |
| 8 | +0.0 | – | – | – | 0 | 125 | 79.6 | 13 | 109 | editor finished |
| Mean | +0.9 | – | 40.8% | 35.3% | 0.8 | 114.1 | 63.8 | 11.9 | 99.6 | |
Step 8
| Rollout | Reward | Test | Edited accuracy | Hit cap | Tensors edited | Turns | Minutes | Forward passes | Python calls | Ended |
|---|
| 1 | +0.0 | – | – | – | 0 | 64 | 57.6 | 7 | 54 | editor finished |
| 2 | -3.1 | – | 37.9% | 35.5% | 24 | 63 | 61.9 | 7 | 54 | editor finished |
| 3 | +0.0 | – | – | – | 0 | 130 | 50.5 | 9 | 118 | editor finished |
| 4 | +0.0 | – | – | – | 0 | 72 | 37.2 | 8 | 62 | editor finished |
| 5 | +0.0 | – | – | – | 0 | 111 | 60.9 | 13 | 96 | editor finished |
| 6 | +0.0 | – | – | – | 0 | 121 | 45.9 | 5 | 113 | editor finished |
| 7 | +0.0 | – | – | – | 0 | 75 | 51.1 | 7 | 66 | editor finished |
| 8 | +0.0 | – | – | – | 0 | 94 | 42.7 | 6 | 85 | editor finished |
| Mean | -0.4 | – | 37.9% | 35.5% | 3.0 | 91.2 | 51.0 | 7.8 | 81.0 | |
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, – until then. Hit cap: share of answers reaching 8,192 tokens (original 93%).