knowledge-commonsense-mix

Fixed evaluation-set composition built with BenchPress. This repo stores only the recipe (manifest.json) — source data is streamed at load time.

Abilities: evidence_retrieval, abstract_pattern_induction

Sources

benchmark	source repo	config	split	n_samples	license	gated
MMLU	tinyBenchmarks/tinyMMLU	all	test	100	unknown	no
ARC Challenge	allenai/ai2_arc	ARC-Challenge	test	150	cc-by-sa-4.0	no
HellaSwag	tinyBenchmarks/tinyHellaswag	default	validation	100	unknown	no

Usage

from benchpress_hub import load_composition

ds = load_composition("seonghyeon0408/knowledge-commonsense-mix")

Sources pinned to commit SHAs; loading is deterministic (fixed items, seeded order).