son of a bitch!!!!
**ALL** of my GRPO runs have been done with an extremely narrow subset of the probability distribution!!!
why is it even fallbacking to that in the first place if vllm accepts `topk = -1` just fine!!!!
>top_k set to -1 -everywhere- in my env code for vllm
>verifiers.envs.rm_env - INFO - top_k: 50
WHERE THE HELL IS THAT BS DEFAULT COMING FROM!!!