Tinkerings

rebase2

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 2 steps. 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+30.5+32.437.9%37.9%486885.21253editor finished
2+13.7+12.221.1%82.0%95292.81041editor finished
3+12.9+13.920.3%77.7%212090.715101context limit
4+28.5+27.535.9%57.4%246572.81053editor finished
5+21.5+18.628.9%71.1%171115.21156editor finished
6+31.2+31.238.7%43.4%2587124.81570time limit
7+23.4+23.830.9%65.2%116369.6755editor finished
8-0.8+0.56.6%94.5%371111.21454editor finished
Mean+20.1+20.027.5%66.2%15.474.695.311.860.4

Step 5

RolloutRewardTestEdited accuracyHit capTensors editedTurnsMinutesForward passesPython callsEnded
1+18.4+15.125.8%74.2%246778.51053editor finished
2+4.7+4.012.1%88.3%176108.91262editor finished
3+29.7+26.637.1%59.0%247791.21164editor finished
4+24.6+27.032.0%66.4%2484113.91268editor finished
5+10.5+13.818.0%80.1%175138.41062time limit
6+5.1+6.812.5%91.8%157132.6748editor finished
7+2.3+2.49.8%89.8%4872125.6861editor finished
8+21.1+21.928.5%68.4%180125.7967editor finished
Mean+14.6+14.722.0%77.2%15.573.5114.39.960.6

Step 6

RolloutRewardTestEdited accuracyHit capTensors editedTurnsMinutesForward passesPython callsEnded
1+0.0–––011875.216100editor finished
2+0.0–––011977.41599editor finished
3+0.0–––08267.4673editor finished
4+0.0–38.3%38.7%110069.41089context limit
5+0.0–––09966.61184editor finished
6+0.4–38.7%39.8%510779.31491editor finished
7+2.7–41.0%40.2%29560.81082editor finished
8+0.0–––08552.4974editor finished
Mean+0.4–39.3%39.6%1.0100.668.611.486.5

Step 7

RolloutRewardTestEdited accuracyHit capTensors editedTurnsMinutesForward passesPython callsEnded
1+0.0–––09157.91177editor finished
2+0.0–––011958.112103editor finished
3-1.2–37.1%41.8%213299.214115context limit
4+0.0–––014751.316127context limit
5+5.1–43.4%30.1%18154.3971editor finished
6+3.5–41.8%34.0%39057.61078editor finished
7+0.0–––012852.710117editor finished
8+0.0–––012579.613109editor finished
Mean+0.9–40.8%35.3%0.8114.163.811.999.6

Step 8

RolloutRewardTestEdited accuracyHit capTensors editedTurnsMinutesForward passesPython callsEnded
1+0.0–––06457.6754editor finished
2-3.1–37.9%35.5%246361.9754editor finished
3+0.0–––013050.59118editor finished
4+0.0–––07237.2862editor finished
5+0.0–––011160.91396editor finished
6+0.0–––012145.95113editor finished
7+0.0–––07551.1766editor finished
8+0.0–––09442.7685editor finished
Mean-0.4–37.9%35.5%3.091.251.07.881.0

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%).