coding-mix

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

Abilities: algorithmic_decomposition, rule_constrained_reasoning

Sources

benchmark	source repo	config	split	n_samples	license	gated
HumanEval	evalplus/humanevalplus	default	test	100	apache-2.0	no
MBPP	google-research-datasets/mbpp	full	test	150	cc-by-4.0	no
LiveCodeBench	livecodebench/code_generation	-	test	100	cc	no

Reference scores (expected on this composition)

model	expected score
Claude-3.5-Sonnet	0.7997
Claude-Sonnet-4	0.8314
DeepSeek-v3	0.604
GPT-4o	0.7206
GPT-oss-20b	0.8069
Gemma-3-4B	0.3671
Llama-3.1-8B	0.5489
Phi-4-Mini	0.5504
Qwen2.5-72B	0.784
Qwen3-235B	0.8086

Usage

from benchpress_hub import load_composition

ds = load_composition("seonghyeon0408/coding-mix")

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