I want both the TaskPrompter and InvPT evaluation packages to expose a stateful…

I want both the TaskPrompter and InvPT evaluation packages to expose a stateful…: a task in MiMo-V2.6-RL-harbor-code: MiMo-V2.6-RL Code (Harbor) (Harbor dataset). Predictions are image-scale saliency values: update should convert them to floating point values divided by 255, evaluate them as…

The task

Predictions are image-scale saliency values: `update` should convert them to floating point values divided by 255, evaluate them as probabilities, ignore target pixels equal to `ignore_index`, and accumulate precision/recall counts over thresholds from `threshold_step` up to but not including 1. If `target` has a…

Part of FineEnvs/MiMo-V2.6-RL-harbor-code.