Chimera Bench: 8,503 Multi-Step Reasoning Problems Across Four Domains

VectionLabs created Chimera Bench V1, a benchmark dataset containing 8,503 articulated multi-step problems. The dataset is organized into four domains: 3,803 problems in MATH, 1,500 in CODE, 1,500 in SCIENCE, and an unspecified number in REASONING. It was last updated on the Hugging Face platform in April 2026.

Use Cases

• Benchmarking large language model performance on multi-step word problems based on the MATH domain's focus.
• Evaluating code reasoning and algorithm design capabilities based on the CODE domain's tasks.
• Testing scientific and physics-based reasoning based on the SCIENCE domain's content.
• Training models for hybrid intelligence tasks based on the dataset's articulated multi-step problem structure.

Strengths

• Contains 8,503 total problems, which is described as larger than the GSM8K benchmark.
• Covers four distinct reasoning domains: MATH, CODE, SCIENCE, and REASONING.
• Includes 3,803 problems specifically in the MATH domain, focusing on multi-step word problems.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count per domain beyond MATH and CODE is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

VectionLabs via Hugging Face.

Collection Method

Likely curated or synthesized for benchmarking purposes.

Time Range

null

Freshness

Last updated 2026-04-15 19:18:01; freshness should be verified.

Geography

null