Synthetic Chain of Reasoning Data for Language Model Training

SCoRe contains 115,659 unique questions and 19,921 DPO pairs structured across 107 reasoning topics and 37 question forms. The dataset was curated from the output distribution of GPT-OSS-120B and Qwen3-32B models, with each record graded and filtered for high-quality reasoning chains. It was created by jon7009 and last updated in March 2026.

Use Cases

• Train preference models using the 19,921 DPO pairs for instruction tuning.
• Analyze reasoning patterns across the 107 distinct topics and 37 question forms.
• Fine-tune language models on high-quality reasoning chains for improved Chain of Thought performance.
• Generate synthetic training data by leveraging the structured curriculum and prompt architecture.

Strengths

• Contains 115,659 unique questions providing substantial training material.
• Systematically covers 107 reasoning topics for broad conceptual coverage.
• Includes 19,921 DPO pairs curated for preference modeling tasks.
• Data is graded, filtered, and postprocessed to retain only high-quality reasoning chains.

Limitations

• Data is synthetic, generated by frontier models, which may introduce biases or artifacts from those sources.
• The dataset's geographic and temporal coverage is unspecified, limiting applicability for certain analyses.
• Lack of provided column details makes direct feature analysis and integration planning difficult.

Provenance

Source

Curated from the output distribution of GPT-OSS 120B and Qwen3 32B models.

Collection Method

Generated using a curriculum and prompt architecture designed with frontier model LLM assistance, then graded, filtered, and postprocessed.

Time Range

null

Freshness

Last updated March 2026.

Geography

null