Tinkerings

rebase1

Branched from nolib1 after step 3; steps 1 to 3 are shared. No notes, no web tools, learning rate 2e-4; the starting model is updated to the best edit so far every step. Updated 13:51 PST, 16 Nov 2026 · all runs

Step 1 (shared with hero3)

RolloutRewardTestEdited accuracyHit capTensors editedTurnsMinutesForward passesPython callsEnded
1+1.2+0.68.6%92.6%176105.61555editor finished
2+9.4+11.116.8%82.4%486652.7754editor finished
3+18.8+16.726.2%78.5%68788.41173editor finished
4+15.6+15.223.0%75.4%67785.91163editor finished
5+33.6+32.841.0%39.8%488466.1972editor finished
6+0.0no edit––0278.6019no tool call in five replies
7+2.7+2.510.2%90.6%196106.21479editor finished
8+27.3+29.434.8%58.6%488659.91469editor finished
Mean+13.6+15.522.9%74.0%19.874.971.710.160.5

Step 2 (shared with nolib1)

RolloutRewardTestEdited accuracyHit capTensors editedTurnsMinutesForward passesPython callsEnded
1+21.1+19.028.5%70.3%488147.1968editor finished
2+11.3+12.318.8%80.5%46973.01254editor finished
3+16.8+18.724.2%74.6%486158.81247editor finished
4+21.1+21.628.5%69.1%19188.81373editor finished
5+0.4+0.47.8%93.0%16380.2950editor finished
6+13.3+10.020.7%83.2%1872119.51353editor finished
7+21.1+21.428.5%69.5%18478.01467editor finished
8+3.1+4.410.5%90.2%486859.0856editor finished
Mean+13.5+13.520.9%78.8%21.173.675.511.258.5

Step 3 (shared with nolib1)

RolloutRewardTestEdited accuracyHit capTensors editedTurnsMinutesForward passesPython callsEnded
1+32.8–40.2%47.7%488666.61369editor finished
2+27.7–35.2%60.2%489785.61581editor finished
3+0.0–––04613.4430no tool call in five replies
4+16.8–24.2%74.6%307351.31060editor finished
5+14.4–21.9%80.5%67868.61066editor finished
6+18.4–25.8%69.9%249789.31578editor finished
7+25.0–32.4%62.5%24105119.11884editor finished
8+14.4–21.9%76.2%488179.81365editor finished
Mean+18.7–28.8%67.4%28.582.971.712.266.6

Step 4

RolloutRewardTestEdited accuracyHit capTensors editedTurnsMinutesForward passesPython callsEnded
1+6.6–14.1%82.8%1072139.01159time limit
2+1.6–9.0%92.2%3100129.21384editor finished
3+8.2–15.6%84.8%4274105.9961editor finished
4+12.1–19.5%81.2%97097.7858editor finished
5+17.6–25.0%72.7%1139140.520115time limit
6+2.3–9.8%92.6%170135.41256time limit
7+16.0–23.4%68.8%771112.91356editor finished
8+12.5–19.9%79.7%1190113.81275editor finished
Mean+9.6–17.0%81.8%10.585.8121.812.270.5

Step 5

RolloutRewardTestEdited accuracyHit capTensors editedTurnsMinutesForward passesPython callsEnded
1+14.4–37.5%56.6%2484119.31369editor finished
2+16.0–39.1%46.5%486698.51052editor finished
3+13.7–36.7%56.6%48095.81266editor finished
4+12.9–35.9%57.8%384108.31071editor finished
5+16.8–39.8%49.6%247865.11164editor finished
6+18.4–41.4%52.3%4119121.91997editor finished
7+10.9–34.0%59.4%487881.61066editor finished
8+16.0–39.1%48.4%66272.11050editor finished
Mean+14.9–37.9%53.4%20.181.495.311.966.9

Step 6

RolloutRewardTestEdited accuracyHit capTensors editedTurnsMinutesForward passesPython callsEnded
1+0.0–––09471.01180editor finished
2+0.0–––0126111.812111editor finished
3+3.1–42.6%49.2%710697.71688editor finished
4+1.2–40.6%48.4%19463.6982editor finished
5+0.0–––011740.6084editor finished
6+0.8–40.2%51.2%1155104.417132context limit
7+1.6–41.0%45.3%410889.01292editor finished
8+0.4–39.8%49.6%2105115.11682editor finished
Mean+0.9–40.9%48.8%1.9113.186.711.693.9

Step 7

RolloutRewardTestEdited accuracyHit capTensors editedTurnsMinutesForward passesPython callsEnded
1-0.4–40.2%48.4%3113106.11395editor finished
2+4.3–44.9%37.5%17293.31257editor finished
3+0.0–––011785.71098editor finished
4-1.6–39.1%51.6%3010226.7063editor finished
5+0.0–40.6%46.5%487105.51073editor finished
6+1.9–42.6%44.5%27695.01361editor finished
7+0.4–41.0%46.5%18771.51371editor finished
8-32.4–8.2%88.3%57926.8057editor finished
Mean-3.5–36.7%51.9%5.891.676.38.971.9

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%).