math-reasoning-mix

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

Abilities: quantitative_manipulation, algorithmic_decomposition

Sources

benchmark	source repo	config	split	n_samples	license	gated
GSM8K	openai/gsm8k	main	test	200	mit	no
MATH-500	HuggingFaceH4/MATH-500	-	test	150	unknown	no
AIME 2024	math-ai/aime24	-	test	30	apache-2.0	no

Reference scores (expected on this composition)

model	expected score
Claude-3.5-Sonnet	0.8249
Claude-Sonnet-4	0.909
DeepSeek-v3	0.857
GLM-4.7	0.9337
GPT-4o	0.8211
GPT-oss-120b	0.9639
GPT-oss-20b	0.9271
Gemma-3-4B	0.3261
KIMI-K2	0.9394
Llama-3.1-8B	0.6016
Phi-4-Mini	0.7576
Qwen2.5-72B	0.8491
Qwen3-235B	0.9513

Usage

from benchpress_hub import load_composition

ds = load_composition("seonghyeon0408/math-reasoning-mix")

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