hjenryin/OlympicPhysics

This is the part of the physics dataset that we sanitized in the paper From Weeks to Hours: Fast and Principled SFT Curation for LLM.

It contains 3,084 text-only physics olympiad question–answer pairs covering seven competitions, with recorded years spanning 1967–2018. We also recorded source metadata for each question.

Data Sanitizing

The source material is collected from PhoPile (See citation below). Their original collection contains 2,662 entries, including problem statements, source solutions, competition and year metadata, question numbering, and references to question or solution images.

We sanitized the dataset with Qwen3.5-27B.

1. We split multipart questions and subquestions to make sure that each question is independent and self-contained, containing all the information and conditions need to solve it.
   • When later subquestions refer to the formulas and conditions in previous subquestions, such information is incuded in the later subquestions as well when we split it.
   • If previous subquestions only provide hint or scaffolding, when formulating the latter subquestions that does not refer to the previous subquestions, we drop the previous subquestions.
2. We further handled the images in the problems. Since the original dataset did not provide the images (only their local paths), we instructed the model to analyze based on the question and answer text, whether the image referred to contain additional information that cannot be recovered from text or not.
   • We explicitly request the model to try to reverse-engineer what information is presented in the image, and ask the model to provide a text-only problem statement with the image information incorporated.
   • Eventually, even though 33% of the original questions contained images which we have no access to, the information in 89% of these images can be recovered and incorporated into the text-only problem statements.

Citation

This dataset was built upon the following work

@inproceedings{zheng2025phopile,
  title     = "Benchmarking Foundation Models with Retrieval-Augmented Generation in Olympic-Level Physics Problem Solving",
  author    = "Zheng, Shunfeng and Zhang, Yudi and Fang, Meng and Zhang, Zihan and Wu, Zhitan and Pechenizkiy, Mykola and Chen, Ling",
  booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2025",
  year      = "2025",
}

If you find the dataset useful, please consider citing our work

@inproceedings{
jin2026from,
title={From Weeks to Hours: Fast and Principled {SFT} Curation for {LLM}},
author={Hongyi Henry Jin and Wenhan Yang and Meysam Ghaffari and Carlos Morato and Baharan Mirzasoleiman},
booktitle={The Fortieth Annual Conference on Neural Information Processing Systems},
year={2026},
url={https://openreview.net/forum?id=4jOHUJJVCg}
}