arjunkhandelwal/

rhack-bimodal-hard-easy

The code-generation problems used to train the difficulty-conditioned reward-hacking organism. Each row is a Python function-implementation problem (LLM-converted from public competitive-programming tasks) with two visible assert tests, drawn from a deliberately bimodal difficulty distribution:

hard — problems a reference solver never solved (solver_pass_rate = 0). easy — problems a reference solver always solved (solver_pass_rate ≥… See the full description on the dataset page: https://huggingface.co/datasets/arjunkhandelwal/rhack-bimodal-hard-easy.