RationaleRM: 10K-100K Samples for Aligning Reward Model Reasoning

RationaleRM provides between 10,000 and 100,000 records designed to align the reasoning processes of reward models with human judgments, released by Qwen in February 2026. The dataset focuses on rationale consistency to distinguish frontier models and detect deceptive alignment in text classification and question-answering tasks. It serves as a benchmark for evaluating whether a model's internal logic matches its final output.

Use Cases

• Training reward models to prioritize rationale-consistency over simple outcome accuracy
• Implementing llm-as-judge workflows to evaluate the logical steps of model responses
• Researching deceptive alignment by analyzing discrepancies between model rationales and final outputs

Strengths

• Scale of 10,000 to 100,000 records
• Includes rationale-consistency metrics for detecting deceptive alignment
• Linked to Arxiv research 260204649 for theoretical grounding

Limitations

• Limited to English language tasks
• Geographic bias toward US-centric data as indicated by region tags
• Lack of detailed column-level documentation in the metadata

Provenance

Source

Qwen (Arxiv 260204649)

Freshness

Last updated February 2026.

Geography

United States