metadata
license: mit
task_categories:
- question-answering
- code-understanding
language:
- en
pretty_name: MiniTorch GraphQA
tags:
- code reasoning
- graph-based QA
- symbolic reasoning
- minitorch
- trm
size_categories:
- n<1K
source_datasets:
- original
MiniTorch GraphQA
A dataset of 100 developer-style questions over the MiniTorch codebase, designed for structured, graph-based code reasoning. Each question maps to a precise function or class in the source code and is intended for use with symbolic reasoning models like the Tiny Recursion Model (TRM).
Intended Use
This dataset is designed to evaluate lightweight, recursive reasoning agents that operate over retrieved subgraphs of code (e.g., 10-node neighborhoods). It is not for code generation or language modeling.
Data Format
Each line in questions.jsonl is a JSON object:
{
"id": "minitorch-001",
"question": "Which function computes the derivative of scalar addition?",
"answer_node": "minitorch.scalar.Scalar.add_back",
"module": 1,
"reasoning_type": "autodiff"
}
id: unique identifier
question: natural language developer question
answer_node: fully qualified name of the ground-truth function/class
module: MiniTorch module number (0–6)
reasoning_type: category of reasoning required
License
MIT License.
Citation
If you use this dataset, please cite:
@dataset{minitorch_graphqa_2025,
author = {Sarosh Quraishi},
title = {MiniTorch GraphQA: A Symbolic Reasoning Benchmark for Code},
year = {2025},
publisher = {Hugging Face},
url = {https://huggingface.co/datasets/saroshq/minitorch-graphqa}
}
Acknowledgements
MiniTorch: https://github.com/atgctg/minitorch