A lightweight problem-source dataset built from aimosprite/high-reasoning-eval.
It contains the 300 hardest problems under the default ranking:
lower pass_at_16 first
then lower majority_at_16 first
then higher disagreement_count_16 first
data/hardest_300.jsonl
manifests/hardest_300_ids.json
manifests/hardness_summary.json
problem_id
problem_key
partition
category
question
answer
selected_answer_hf… See the full description on the dataset page:
https://huggingface.co/datasets/aimosprite/high-reasoning-clean-eval.
