Datasets:

vectionlabs
/

chimera-bench-v1

language:
  - en
license: apache-2.0
task_categories:
  - text-generation
  - question-answering
tags:
  - benchmark
  - coding
  - math
  - science
  - reasoning
  - logic
  - evaluation
pretty_name: CHIMERA Bench v3.0 Mega
size_categories:
  - 1K<n<10K

CHIMERA Bench v3.0 Mega

Comprehensive Hybrid Intelligence Metric for Excellence in Reasoning & Analysis

8503 articulated multi-step problems across 4 domains (larger than GSM8K).

Domain	Problems	Focus
MATH	3803	Multi-step word problems: shopping, speed/distance, geometry, combinatorics, algebra, number theory, calculus
CODE	1500	Code tracing, bug finding, algorithm design, complexity analysis, OOP, recursion
SCIENCE	1500	Physics (projectile, energy, circuits), chemistry (stoichiometry, pH, gas laws), biology (genetics, ecology)
THINK	1700	Constraint satisfaction, knights & knaves, scheduling, deduction, pattern recognition, estimation
Total	8503

Usage

from datasets import load_dataset
ds = load_dataset("vectionlabs/chimera-bench-v1", split="test")
print(f"{len(ds)} problems loaded")
for p in ds.select(range(5)):
    print(f"[{p['domain']}] {p['title']}: {p['prompt'][:80]}...")

Evaluation Types

Type	Method
`numeric`	Value within tolerance
`exact`	Normalized string match
`keyword`	Required keywords present

Difficulty: 1-5 stars (10-100 points)

Citation

@misc{chimera-bench-v3-2025,
  title={CHIMERA Bench v3: Comprehensive Hybrid Intelligence Metric for Excellence in Reasoning and Analysis},
  year={2025},
  url={https://huggingface.co/datasets/vectionlabs/chimera-bench-v1}
}