narabzad/t3-struct-qwq32b

T3: Transformation of Thinking Traces (QwQ-32B)

This dataset contains T3-transformed thinking traces (Structural Normalization) derived from the QwQ-32B model, as introduced in the paper RAG over Thinking Traces Can Improve Reasoning Tasks.

Retrieval-augmented generation (RAG) is traditionally believed to offer limited benefit for reasoning-intensive tasks like math and code. This work challenges that assumption by showing that retrieving thinking traces—intermediate reasoning trajectories from strong models—consistently improves performance across frontier models and benchmarks. T3 is an offline method that transforms raw traces into structured, retrieval-friendly representations.

Links

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

Sample Usage

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

from datasets import load_dataset

# Load the T3-transformed passages
ds = load_dataset("narabzad/t3-struct-qwq32b")
# Columns: question, trace, passages (list)

Citation

@article{arabzadeh2026rag,
  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={2026}
}