Add nvidia/OpenMathReasoning dataset (cot, tir, genselect splits)

[XinyuGuan]

Description

Fixes #1848(Partially)

This PR adds the nvidia/OpenMathReasoning dataset to Marin's SFT dataset collection. OpenMathReasoning is a high-quality mathematical reasoning dataset containing 5.67 million samples generated by DeepSeek-R1 and QwQ-32B models (Qwen2.5-32B-Instruct to preprocess problems, and DeepSeek-R1 and QwQ-32B to generate solutions), featuring chain-of-thought solutions to olympiad-level math problems.

Dataset Details

• Source: https://huggingface.co/datasets/nvidia/OpenMathReasoning
• License: CC-BY-4.0
• Revision: d3d0866

Splits Added

Split	Samples	Description
cot	3.2M	Chain-of-thought reasoning solutions
tir	1.72M	Tool-integrated reasoning (code execution)
genselect	566K	Generated selection samples

Field Mapping

• problem → instruction column
• generated_solution → response column
• Metadata columns preserved: expected_answer, problem_source, problem_type, generation_model, inference_mode, pass_rate_72b_tir, used_in_kaggle

Testing Performed

• Pre-commit checks passed
• pytest: 1 passed, 3 skipped, 19 deselected
• Verified config loads correctly and matches structure of existing configs (e.g., meta-math/MetaMathQA)
• Verified dataset field names (problem, generated_solution) exist via HuggingFace streaming
• End-to-end SFT training test with 5000 samples on 8×A800 GPUs (from distilgpt2 - 82M):
  • Loss decreased from 2.87 → 1.52 over 15 epochs
  • Model successfully learned the CoT reasoning format from training data
• Example:

  • Question: How many lines can be drawn that are equidistant from 3 non-collinear points in the plane?

  • Before SFT:

    Click to view output

    The solution is to use the plane geometry, then fill the plane.
    #include <math.h> #include <std.h> #include <math.h> #include <math.h> #include <math.h> #include <math.h> #include <math.h> #include <math.h> #include <math.h> #include <math.h> #include <math.h> #include <math.h> #include <math.h> #include <math.h> #include <math.h> #include <math.h> #include <math.h> #include <math.h> #include <math.h> #include <math.h> #include <math.h> #include <math.h> #include <math.h> #include <math.h> #include <math.h> #include <math.h> #include <math.h> #include <math.h> #include <math.h> #include <math.h> #include <math.h> #include <math.h> #include <math.h> #include <math.h> #include <math.h> #include <math.h> #include <math.h> #include <math.h> #include <math.h> #include <math.h> #include <math.
    

  • After SFT:

    Click to view output

    <think>
    Okay, so I need to figure out how many lines are equidistant from three non-collinear points in the plane. Hmm, let's see. The problem is asking for the number of lines that can be drawn that are equidistant from 3 non-collinear points in the plane. The condition is that the line is equidistant from all three non-collinear points in that plane.
    
    First, maybe I should recall some properties of line-drawing. Since all lines have at least one intersection point, but the problem is asking for the number of lines that aren't equidistant at all, maybe there's a pattern or a way to relate these lines to other points.
    
    Since all lines are equidistant from all three non-collinear points, maybe the line that's equidistant from all three non-collinear points in that plane is equidistant from all three non-collinear points in that plane. Since all lines are equidistant from all three non-collinear points, that means all three non-collinear points must be equidistant from all three non-collinear points in that plane.
    
    I remember that when dealing with lines equidistant from a plane, the intersection points of the sides of two lines are the plane's transversals, and the line that's equidistant from all three non-
    

Checklist

• [ Passed ] You ran uv run python infra/pre-commit.py --all-files to lint/format your code
• [ 1 passed, 3 skipped, 19 deselected ] You ran 'pytest' to test your code

[dlwh]

thanks so much for your contribution! This looks good to me!

Double checking that you meant to leave out additional_problems

[XinyuGuan]

thanks so much for your contribution! This looks good to me!

Double checking that you meant to leave out additional_problems

Hi,

Thank you for your comment! The additional 193K problems don't have solutions (labels) so I didn't add them. If you think they are useful, I can add the split next week. I can also add other datasets mentioned in #1848.