CreativeBench: 1,859 Tasks for Evaluating Code Model Creativity

CreativeBench contains 1,859 JSONL records developed by Zethive in 2026 to evaluate the creative problem-solving capabilities of code models. The data is partitioned into two subsets focusing on combinatorial creativity (1,308 records) and exploratory creativity (551 records).

Use Cases

• Evaluating code model performance on combinatorial creativity tasks using the 1,308-record subset
• Benchmarking exploratory problem-solving logic in LLMs using the 551-record subset
• Testing novel code generation capabilities for non-standard programming challenges

Strengths

• 1,859 total records
• Dual-subset structure for combinatorial and exploratory creativity
• JSONL format for stream processing

Limitations

• Small total sample size of 1,859 records
• Unknown column schema in metadata

Provenance

Source

Zethive via HuggingFace

Freshness

Last updated March 2026