Tri-NL: 1,500 Multi-Domain Reasoning Samples in Math, Coding, and Healthcare

Tri-NL provides 1,500 synthetic reasoning samples across mathematics, coding, and healthcare domains, published by crevious in March 2026. Each record contains a problem and a solution structured into numbered reasoning steps and a final answer.

Use Cases

• Fine-tuning models for chain-of-thought reasoning using the 'Step 1:' to 'Final Answer:' formatting
• Evaluating cross-domain logic by testing models on the 15% biostatistics overlap between healthcare and math
• Benchmarking domain-specific accuracy within the 500-sample coding subset

Strengths

• 1,500 samples with 500 per domain
• Standardized 'Step 1:' reasoning format
• 15% intentional cross-domain overlap

Limitations

• Small sample size of 1,500 records
• Synthetic nature may lack real-world edge cases
• Unknown column names

Provenance

Source

crevious

Collection Method

synthetic

Freshness

Last updated March 2026.