narabzad/t3-traces-qwq32b

T3: Raw Thinking Traces (GPT-OSS-120B)

This dataset contains raw thinking traces—intermediate reasoning trajectories—generated by the GPT-OSS-120B model. These traces were produced during problem-solving attempts across various reasoning-intensive benchmarks.

This work was introduced in the paper: RAG over Thinking Traces Can Improve Reasoning Tasks.

The authors propose that Retrieval-Augmented Generation (RAG) is highly effective for reasoning tasks (like math and code) when the retrieval corpus consists of thinking traces rather than standard web documents. These traces provide a procedural scaffold that helps downstream models solve complex problems without additional fine-tuning.

• —Paper: RAG over Thinking Traces Can Improve Reasoning Tasks
• —Repository: https://github.com/Narabzad/t3

Sample Usage

You can load the dataset using the Hugging Face datasets library:

from datasets import load_dataset

# Load raw thinking traces for GPT-OSS-120B
ds = load_dataset("narabzad/t3-traces-gptoss120b")

# The dataset contains two columns:
# - question: The input problem/query
# - trace: The model's intermediate reasoning trajectory
print(ds['train'][0])

Dataset Structure

The dataset consists of a single train split with the following columns:

• —question: string - The original problem or query.
• —trace: string - The raw intermediate thinking trajectory generated by the model.

Citation

@article{arabzadeh2024rag,
  title={RAG over Thinking Traces Can Improve Reasoning Tasks},
  author={Arabzadeh, Negar and Ma, Wenjie and Min, Sewon and Zaharia, Matei},
  journal={arXiv preprint arXiv:2605.03344},
  year={2024}
}