TRM-Preference: Reasoning Trace Quality for Large Reasoning Models

TRM-Preference is a dataset introduced in the paper 'Characterizing, Evaluating, and Optimizing Complex Reasoning' for evaluating and optimizing reasoning traces in Large Reasoning Models. The dataset applies the ME² principle to assess 'how a model thinks' across dimensions like Macro-Efficiency. It was authored by zzzhr97 and last updated on Hugging Face in June 2026.

Use Cases

• Training a Thinking Reward Model (TRM) based on the described ME² evaluation principle.
• Evaluating the quality of reasoning traces in Large Reasoning Models beyond simple answer correctness.
• Optimizing model reasoning processes based on multi-dimensional criteria like Macro-Efficiency.

Strengths

• Dataset is designed for a specific, published research purpose as described in the paper 'Characterizing, Evaluating, and Optimizing Complex Reasoning'.
• Focuses on evaluating reasoning quality across multiple dimensions, not just final answer correctness.

Limitations

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

Provenance

Source

zzzhr97 on Hugging Face.

Collection Method

Likely created for the research paper 'Characterizing, Evaluating, and Optimizing Complex Reasoning'.

Freshness

Last updated 2026-06-05 13:40:02; freshness should be verified.