lime-nlp/orz_math_difficulty
Difficulty Estimation on Open Reasoner Zero We annotate the entire Open Reasoner Zero dataset with a difficulty score based on the performance of the Qwen 2.5-MATH-7B model. This provides an adaptive signal for curriculum construction. Open Reasoner Zero is a curated a dataset of 57,000 reasoning-intensive problems used to train and evaluate reinforcement learning-based methods for large language models. Difficulty Scoring Method Difficulty scores are estimated… See the full description on the dataset page: https://huggingface.co/datasets/lime-nlp/orz_math_difficulty.
042
1---2license: mit3configs:4- config_name: Difficulty Score5 data_files: Qwen2.5-Math-7B--orz--difficulty.csv6- config_name: Response7 data_files: Qwen2.5-Math-7B--orz.csv8---9 10## Difficulty Estimation on Open Reasoner Zero11We annotate the entire [**Open Reasoner Zero**]((https://huggingface.co/Open-Reasoner-Zero/Open-Reasoner-Zero-7B)) dataset with a **difficulty score** based on the performance of the [Qwen 2.5-MATH-7B](https://huggingface.co/Qwen/Qwen2.5-Math-7B) model. This provides an adaptive signal for curriculum construction.12Open Reasoner Zero is a curated a dataset of 57,000 reasoning-intensive problems used to train and evaluate reinforcement learning-based methods for large language models.13 14## Difficulty Scoring Method15 16Difficulty scores are estimated using the **Qwen 2.5-MATH-7B** model with the following generation settings:17 18- `temperature = 0.6`19- `top_p = 0.9`20- `max_tokens = 4096`21- Inference performed using [vLLM](https://github.com/vllm-project/vllm)22- Each problem is attempted **128 times**23 24The difficulty score `d_i` for each problem is computed as:25 26 d_i = 100 × (1 - (# successes / 128))27 28This approach balances the evaluation signal:29- A **strong model** would trivially solve easy problems, compressing the difficulty scale.30- A **weak model** would fail uniformly, providing poor resolution.31- Qwen 2.5-MATH-7B was selected for its **mid-range capabilities**, offering meaningful gradients across a wide spectrum of problems.32 33## Difficulty Estimation on Other Datasets34 35We also apply the same difficulty estimation procedure to the following datasets:36 37- [Open Reasoner Zero](https://huggingface.co/datasets/lime-nlp/orz_math_difficulty)38- [DeepScaleR](https://huggingface.co/datasets/lime-nlp/DeepScaleR_Difficulty)39- [MATH](https://huggingface.co/datasets/lime-nlp/MATH_difficulty)40- [GSM8K](https://huggingface.co/datasets/lime-nlp/GSM8K_difficulty)41 42## 📬 Contact43 44For questions or feedback, feel free to reach out to [**Taiwei Shi**](https://maksimstw.github.io/) at [taiweish@usc.edu](mailto:taiweish@usc.edu).45 46## 📚 Citations47Github: https://github.com/uscnlp-lime/verl48 49If you find our dataset useful, please cite [Efficient Reinforcement Finetuning via Adaptive Curriculum Learning](https://huggingface.co/papers/2504.05520):50 51```bibtex52@misc{shi2025efficientreinforcementfinetuningadaptive,53 title={Efficient Reinforcement Finetuning via Adaptive Curriculum Learning}, 54 author={Taiwei Shi and Yiyang Wu and Linxin Song and Tianyi Zhou and Jieyu Zhao},55 year={2025},56 eprint={2504.05520},57 archivePrefix={arXiv},58 primaryClass={cs.LG},59 url={https://arxiv.org/abs/2504.05520}, 60}61```