TIRE

# TIRE-PHYSICS-2000 **English** | [简体中文](README_zh.md) Hugging Face: [ScienceOne-AI/TIRE](https://huggingface.co/datasets/ScienceOne-AI/TIRE) GitHub: [ScienceOne-AI/TIRE](https://github.com/ScienceOne-AI/TIRE) Paper: [TIRE: Top-Down Formula Instantiation with Reliable Execution for Physics Problem Synthesis](https://arxiv.org/abs/2610.25134) ## Dataset Overview TIRE-PHYSICS-2000 is a 2,000-record English physics reasoning dataset from the [TIRE](https://github.com/ScienceOne-AI/TIRE) project. It is a curated subset for studying formula-guided physics problem construction and reasoning-data quality. Each record contains one problem and its teacher response. ## Dataset Composition | Item | Value | | --- | --- | | Records | 2,000 | | Language | English | | Source | PHYSICS | | Selection seed | 42 | | Source subject labels | 10 | | Normalized subject groups | 6 | | Distinct formula IDs | 1,953 | The dataset covers the source subjects and formula groups available during selection. Formula IDs and source problems are coverage proxies; they are not an exhaustive taxonomy of physics knowledge. ## Data Fields Each line of train.jsonl is a JSON object with exactly two string fields: | Field | Type | Description | | --- | --- | --- | | question | string | English physics problem and answer-format instruction | | answer | string | Stored teacher reasoning and final response | The final response after </think> contains one complete boxed answer. The reasoning block is retained as stored and may contain intermediate boxed expressions. Example: ~~~json { "question": "A physical system ... put your final answer within \\boxed{}.", "answer": "<think>...reasoning...</think>\n\nTherefore, ... \\boxed{...}" } ~~~ ## Load the Dataset From Hugging Face: ~~~python from datasets import load_dataset dataset = load_dataset("ScienceOne-AI/TIRE", split="train") print(dataset) ~~~ From the local file: ~~~python dataset = load_dataset("json", data_files="train.jsonl", split="train") ~~~ ## License The dataset is released under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/). The generation pipeline is available in [ScienceOne-AI/TIRE](https://github.com/ScienceOne-AI/TIRE) under Apache-2.0. ## 📖 Citation If you use this dataset or the generation pipeline, please cite our paper. ```bibtex @misc{tire2026, title = {TIRE: Top-Down Formula Instantiation with Reliable Execution for Physics Problem Synthesis}, author = {Xiaoyu Wu and Zhaoxin Yu and Xinglin Xiao and Nan Xu}, year = {2026}, url = {https://arxiv.org/abs/2610.25134} } ```