rebase1
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 | +6.6 | – | 14.1% | 82.8% | 10 | 72 | 139.0 | 11 | 59 | time limit |
| 2 | +1.6 | – | 9.0% | 92.2% | 3 | 100 | 129.2 | 13 | 84 | editor finished |
| 3 | +8.2 | – | 15.6% | 84.8% | 42 | 74 | 105.9 | 9 | 61 | editor finished |
| 4 | +12.1 | – | 19.5% | 81.2% | 9 | 70 | 97.7 | 8 | 58 | editor finished |
| 5 | +17.6 | – | 25.0% | 72.7% | 1 | 139 | 140.5 | 20 | 115 | time limit |
| 6 | +2.3 | – | 9.8% | 92.6% | 1 | 70 | 135.4 | 12 | 56 | time limit |
| 7 | +16.0 | – | 23.4% | 68.8% | 7 | 71 | 112.9 | 13 | 56 | editor finished |
| 8 | +12.5 | – | 19.9% | 79.7% | 11 | 90 | 113.8 | 12 | 75 | editor finished |
| Mean | +9.6 | – | 17.0% | 81.8% | 10.5 | 85.8 | 121.8 | 12.2 | 70.5 | |
Step 5
| Rollout | Reward | Test | Edited accuracy | Hit cap | Tensors edited | Turns | Minutes | Forward passes | Python calls | Ended |
|---|
| 1 | +14.4 | – | 37.5% | 56.6% | 24 | 84 | 119.3 | 13 | 69 | editor finished |
| 2 | +16.0 | – | 39.1% | 46.5% | 48 | 66 | 98.5 | 10 | 52 | editor finished |
| 3 | +13.7 | – | 36.7% | 56.6% | 4 | 80 | 95.8 | 12 | 66 | editor finished |
| 4 | +12.9 | – | 35.9% | 57.8% | 3 | 84 | 108.3 | 10 | 71 | editor finished |
| 5 | +16.8 | – | 39.8% | 49.6% | 24 | 78 | 65.1 | 11 | 64 | editor finished |
| 6 | +18.4 | – | 41.4% | 52.3% | 4 | 119 | 121.9 | 19 | 97 | editor finished |
| 7 | +10.9 | – | 34.0% | 59.4% | 48 | 78 | 81.6 | 10 | 66 | editor finished |
| 8 | +16.0 | – | 39.1% | 48.4% | 6 | 62 | 72.1 | 10 | 50 | editor finished |
| Mean | +14.9 | – | 37.9% | 53.4% | 20.1 | 81.4 | 95.3 | 11.9 | 66.9 | |
Step 6
| Rollout | Reward | Test | Edited accuracy | Hit cap | Tensors edited | Turns | Minutes | Forward passes | Python calls | Ended |
|---|
| 1 | +0.0 | – | – | – | 0 | 94 | 71.0 | 11 | 80 | editor finished |
| 2 | +0.0 | – | – | – | 0 | 126 | 111.8 | 12 | 111 | editor finished |
| 3 | +3.1 | – | 42.6% | 49.2% | 7 | 106 | 97.7 | 16 | 88 | editor finished |
| 4 | +1.2 | – | 40.6% | 48.4% | 1 | 94 | 63.6 | 9 | 82 | editor finished |
| 5 | +0.0 | – | – | – | 0 | 117 | 40.6 | 0 | 84 | editor finished |
| 6 | +0.8 | – | 40.2% | 51.2% | 1 | 155 | 104.4 | 17 | 132 | context limit |
| 7 | +1.6 | – | 41.0% | 45.3% | 4 | 108 | 89.0 | 12 | 92 | editor finished |
| 8 | +0.4 | – | 39.8% | 49.6% | 2 | 105 | 115.1 | 16 | 82 | editor finished |
| Mean | +0.9 | – | 40.9% | 48.8% | 1.9 | 113.1 | 86.7 | 11.6 | 93.9 | |
Step 7
| Rollout | Reward | Test | Edited accuracy | Hit cap | Tensors edited | Turns | Minutes | Forward passes | Python calls | Ended |
|---|
| 1 | -0.4 | – | 40.2% | 48.4% | 3 | 113 | 106.1 | 13 | 95 | editor finished |
| 2 | +4.3 | – | 44.9% | 37.5% | 1 | 72 | 93.3 | 12 | 57 | editor finished |
| 3 | +0.0 | – | – | – | 0 | 117 | 85.7 | 10 | 98 | editor finished |
| 4 | -1.6 | – | 39.1% | 51.6% | 30 | 102 | 26.7 | 0 | 63 | editor finished |
| 5 | +0.0 | – | 40.6% | 46.5% | 4 | 87 | 105.5 | 10 | 73 | editor finished |
| 6 | +1.9 | – | 42.6% | 44.5% | 2 | 76 | 95.0 | 13 | 61 | editor finished |
| 7 | +0.4 | – | 41.0% | 46.5% | 1 | 87 | 71.5 | 13 | 71 | editor finished |
| 8 | -32.4 | – | 8.2% | 88.3% | 5 | 79 | 26.8 | 0 | 57 | editor finished |
| Mean | -3.5 | – | 36.7% | 51.9% | 5.8 | 91.6 | 76.3 | 8.9 | 71.9 | |
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%).