VITHURSHAN/DFO-Kaggle

Motivation

Recently, NVIDIA released the OpenMathReasoning dataset. While the main dataset includes model-generated responses for CoT (Chain-of-Thought) and TIR (Tool-Integrated Reasoning), the additional split (~193K problems) contains only questions, with no model-generated responses.

At the same time, it is difficult to find high-quality datasets suitable for Direct Preference Optimization (DPO) in mathematical reasoning—especially datasets that include tool-calling behavior and explicit accepted/rejected preference pairs.

To address these gaps, this dataset provides accepted and rejected responses generated using GPT-OSS-120B with tool-integrated reasoning, explicitly designed for preference-based training (DPO/GRPO/RLHF-style) in mathematical reasoning tasks.

---

Data Source

• NVIDIA OpenMathReasoning – Additional Set

• Original split contains ~193K questions with no model responses

• This processed dataset contains:

  • >4,000 unique questions
  • ~48,000 unique accepted–rejected preference pairs

---

Data Generation & Preparation

Response Generation

For each question:

• Model: gpt-oss-120b
• Inference mode: Tool-Integrated Reasoning (TIR)
• Number of generations: 8 responses per question
• Temperature: 1.0
• min_p: 0.02
• Python tool calling: Enabled

---

Acceptance / Rejection Criteria

From the 8 generated responses:

• Accepted

  • Final answer is correct
  • Zero Python tool execution errors

• Rejected

  • Final answer is incorrect, or
  • Final answer is correct but has one or more Python tool execution errors

---

Preference Pair Construction

For each question, all unique accepted–rejected combinations were constructed in a structured manner:

pairs = []
for acc in accepted_responses:
    for rej in rejected_responses:
        pairs.append({
            "accepted": acc,
            "rejected": rej
        })

This results in multiple high-quality preference pairs per question, suitable for DPO-style optimization.

---

Pass Rate Calculation

The pass_rate is computed as:

pass_rate = (number_of_accepted_responses / 8) * 100

This provides a useful signal for problem difficulty estimation and model performance analysis.

---

Dataset Fields

Each data entry contains the following fields:

{
  'id': id,
  'question': question,
  'generation_model': 'gpt-oss-120b',
  'inference_mode': 'tir',
  'pass_rate': pass_rate,
  'expected_answer': expected_answer,
  'accepted_formatted': accepted_formatted,
  'rejected_formatted': rejected_formatted,
  'license': 'cc-by-4.0'
}

• The GPT-OSS chat template is applied consistently to all responses.
• Accepted and rejected outputs preserve tool calls, reasoning steps, and formatting.

---

License

• CC-BY-4.0