openo1-sft-ultra-35m-data

Instruction

We have released the openo1-sft-ultra-35m-data, which contains 35 million data points. It is based on existing open-source datasets and synthesized using the openo1-qwen-sft model. We first collected open-source datasets and then annotated the data based on difficulty, quality, and question types using the qwen-2.5-72b-instruct model. To ensure the difficulty and quality of the data, we only retained data where both the difficulty and quality are ≥8.

Data format

• 'uid': Data ID
• 'query': Original data query
• 'response': Long COT response, including detailed thought process
• 'source': Data source
• 'difficulty': Question difficulty, range from 1 to 10
• 'quality': Data quality, range from 1 to 10
• 'answer': Ground truth answer of the data
• 'query_len': Length of the question
• 'response_len': Length of the answer
• 'Topic': Data topic category, including Math | Code | Reasoning
• 'answer_type': Answer type annotation
  • For math domain:
    • a: Purely numerical answer
    • b: Purely formulaic answer
    • c: Long textual explanation
  • For code domain:
    • a: Answer that includes code
    • b: Answer that contains only code-related text
  • For reasoning domain:
    • a: Answer consisting of only words/phrases
    • b: Answer that includes a long textual explanation
• "cases": Test cases included in the code

Data statistics

Data source

We have used the following sources for the data:

• WebInstructFull
• homework
• infinity-instruct
• math-stack-exchange
• MathInstruct
• mcq