Answers get shorter as the model learns
Mean answer length in tokens per step for three seeds. The cap is 1,024 tokens. Answers start near 720 tokens and end near 460, so fewer answers hit the cap.
400
500
600
700
800
0
12
24
36
47
step
mean answer length, tokens
seed 1234
seed 2
seed 3
Percent of answers cut at the cap: 33 to 8; 30 to 12; 34 to 9.