Hi authors,
I had a quick question about reproducing your results for the Qwen3-4B distilling Qwen3-1.7B setup. I'm trying to pin down the exact values for NUM_ROLLOUT and ROLLOUT_MAX_RESPONSE_LEN. Some of the .sh files in your repo set NUM_ROLLOUT=50 and ROLLOUT_MAX_RESPONSE_LEN=64, but when I run with those, all the methods (entropy-based, TIP, and TA-OPD) end up underperforming the naive OPD baseline on AIME2024. I did manage to reproduce that baseline exactly though (0.10). For reference, I have the token selection ratio set to 10%.
It's totally possible I'm missing something on my end. Aside from NUM_ROLLOUT and ROLLOUT_MAX_RESPONSE_LEN, are there any other settings or details I might be overlooking? Thanks so much for the help!
Best,
George
Hi authors,
I had a quick question about reproducing your results for the Qwen3-4B distilling Qwen3-1.7B setup. I'm trying to pin down the exact values for NUM_ROLLOUT and ROLLOUT_MAX_RESPONSE_LEN. Some of the .sh files in your repo set NUM_ROLLOUT=50 and ROLLOUT_MAX_RESPONSE_LEN=64, but when I run with those, all the methods (entropy-based, TIP, and TA-OPD) end up underperforming the naive OPD baseline on AIME2024. I did manage to reproduce that baseline exactly though (0.10). For reference, I have the token selection ratio set to 10%.
It's totally possible I'm missing something on my end. Aside from NUM_ROLLOUT and ROLLOUT_MAX_RESPONSE_LEN, are there any other settings or details I might be overlooking? Thanks so much for the help!
Best,
George