talon-community/CodeDebugReasoning

---
license: cc-by-4.0
task_categories:
- question-answering
language:
- en
tags:
- code
- programming
- debug
- reasoning
- thinking
- distill
size_categories:
- 1K<n<10K
---
OpenDebugReasoning is a small, focused dataset for evaluating code debugging and reasoning ability in language models. It contains 1,000 samples derived from ```Vezora/Open-Critic-GPT```, seeded and filtered for quality, then annotated using Gemini API completions.

Each entry in the dataset includes a buggy code snippet and a prompt asking an AI model to identify and fix the issue. The dataset also includes step-by-step reasoning generated during the debugging process.

## Columns
- prompt: A randomly selected natural language instruction asking the model to identify or fix a bug. Ten distinct prompts are used, chosen at random across the dataset.
- code: The buggy code snippet to be analyzed and corrected.
- ground_truth: The corrected version of the code.
- thinking_trajectory: A chain-of-thought style reasoning path showing how the model arrived at a solution.

## Splits
The dataset is divided into two distinct evaluation splits:

- Reference Split: The model is shown the buggy code and the ground truth. This setup is useful for controlled reasoning tasks and interpretability studies.
- Independent Split (Recommended): The model is shown only the prompt and buggy code. It must reason independently and produce the fixed code from scratch. This setup better reflects real-world generalization and debugging skill.