RL training on small catalog Environments no longer needs task counts
rl.train("nodus/[email protected]") and rl.grpo_lora(..., environment="nodus/[email protected]")
were refused with “the train split has 64 tasks” (or 200), because the Python SDK asked for 256 training tasks by
default. The defaults are now 256 training and 64 held-out tasks, or the Environment’s own split sizes when those are
smaller. Counts you pass yourself are sent as given.