Enregistré dans:
| Auteurs principaux: | , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , |
|---|---|
| Format: | Preprint |
| Publié: |
2025
|
| Sujets: | |
| Accès en ligne: | https://arxiv.org/abs/2506.04178 |
| Tags: |
Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
|
| _version_ | 1866916780303187968 |
|---|---|
| author | Guha, Etash Marten, Ryan Keh, Sedrick Raoof, Negin Smyrnis, Georgios Bansal, Hritik Nezhurina, Marianna Mercat, Jean Vu, Trung Sprague, Zayne Suvarna, Ashima Feuer, Benjamin Chen, Liangyu Khan, Zaid Frankel, Eric Grover, Sachin Choi, Caroline Muennighoff, Niklas Su, Shiye Zhao, Wanjia Yang, John Pimpalgaonkar, Shreyas Sharma, Kartik Ji, Charlie Cheng-Jie Deng, Yichuan Pratt, Sarah Ramanujan, Vivek Saad-Falcon, Jon Li, Jeffrey Dave, Achal Albalak, Alon Arora, Kushal Wulfe, Blake Hegde, Chinmay Durrett, Greg Oh, Sewoong Bansal, Mohit Gabriel, Saadia Grover, Aditya Chang, Kai-Wei Shankar, Vaishaal Gokaslan, Aaron Merrill, Mike A. Hashimoto, Tatsunori Choi, Yejin Jitsev, Jenia Heckel, Reinhard Sathiamoorthy, Maheswaran Dimakis, Alexandros G. Schmidt, Ludwig |
| author_facet | Guha, Etash Marten, Ryan Keh, Sedrick Raoof, Negin Smyrnis, Georgios Bansal, Hritik Nezhurina, Marianna Mercat, Jean Vu, Trung Sprague, Zayne Suvarna, Ashima Feuer, Benjamin Chen, Liangyu Khan, Zaid Frankel, Eric Grover, Sachin Choi, Caroline Muennighoff, Niklas Su, Shiye Zhao, Wanjia Yang, John Pimpalgaonkar, Shreyas Sharma, Kartik Ji, Charlie Cheng-Jie Deng, Yichuan Pratt, Sarah Ramanujan, Vivek Saad-Falcon, Jon Li, Jeffrey Dave, Achal Albalak, Alon Arora, Kushal Wulfe, Blake Hegde, Chinmay Durrett, Greg Oh, Sewoong Bansal, Mohit Gabriel, Saadia Grover, Aditya Chang, Kai-Wei Shankar, Vaishaal Gokaslan, Aaron Merrill, Mike A. Hashimoto, Tatsunori Choi, Yejin Jitsev, Jenia Heckel, Reinhard Sathiamoorthy, Maheswaran Dimakis, Alexandros G. Schmidt, Ludwig |
| contents | Reasoning models have made rapid progress on many benchmarks involving math, code, and science. Yet, there are still many open questions about the best training recipes for reasoning since state-of-the-art models often rely on proprietary datasets with little to no public information available. To address this, the goal of the OpenThoughts project is to create open-source datasets for training reasoning models. After initial explorations, our OpenThoughts2-1M dataset led to OpenThinker2-32B, the first model trained on public reasoning data to match DeepSeek-R1-Distill-32B on standard reasoning benchmarks such as AIME and LiveCodeBench. We then improve our dataset further by systematically investigating each step of our data generation pipeline with 1,000+ controlled experiments, which led to OpenThoughts3. Scaling the pipeline to 1.2M examples and using QwQ-32B as teacher yields our OpenThoughts3-7B model, which achieves state-of-the-art results: 53% on AIME 2025, 51% on LiveCodeBench 06/24-01/25, and 54% on GPQA Diamond - improvements of 15.3, 17.2, and 20.5 percentage points compared to the DeepSeek-R1-Distill-Qwen-7B. All of our datasets and models are available on https://openthoughts.ai. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2506_04178 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | OpenThoughts: Data Recipes for Reasoning Models Guha, Etash Marten, Ryan Keh, Sedrick Raoof, Negin Smyrnis, Georgios Bansal, Hritik Nezhurina, Marianna Mercat, Jean Vu, Trung Sprague, Zayne Suvarna, Ashima Feuer, Benjamin Chen, Liangyu Khan, Zaid Frankel, Eric Grover, Sachin Choi, Caroline Muennighoff, Niklas Su, Shiye Zhao, Wanjia Yang, John Pimpalgaonkar, Shreyas Sharma, Kartik Ji, Charlie Cheng-Jie Deng, Yichuan Pratt, Sarah Ramanujan, Vivek Saad-Falcon, Jon Li, Jeffrey Dave, Achal Albalak, Alon Arora, Kushal Wulfe, Blake Hegde, Chinmay Durrett, Greg Oh, Sewoong Bansal, Mohit Gabriel, Saadia Grover, Aditya Chang, Kai-Wei Shankar, Vaishaal Gokaslan, Aaron Merrill, Mike A. Hashimoto, Tatsunori Choi, Yejin Jitsev, Jenia Heckel, Reinhard Sathiamoorthy, Maheswaran Dimakis, Alexandros G. Schmidt, Ludwig Machine Learning Reasoning models have made rapid progress on many benchmarks involving math, code, and science. Yet, there are still many open questions about the best training recipes for reasoning since state-of-the-art models often rely on proprietary datasets with little to no public information available. To address this, the goal of the OpenThoughts project is to create open-source datasets for training reasoning models. After initial explorations, our OpenThoughts2-1M dataset led to OpenThinker2-32B, the first model trained on public reasoning data to match DeepSeek-R1-Distill-32B on standard reasoning benchmarks such as AIME and LiveCodeBench. We then improve our dataset further by systematically investigating each step of our data generation pipeline with 1,000+ controlled experiments, which led to OpenThoughts3. Scaling the pipeline to 1.2M examples and using QwQ-32B as teacher yields our OpenThoughts3-7B model, which achieves state-of-the-art results: 53% on AIME 2025, 51% on LiveCodeBench 06/24-01/25, and 54% on GPQA Diamond - improvements of 15.3, 17.2, and 20.5 percentage points compared to the DeepSeek-R1-Distill-Qwen-7B. All of our datasets and models are available on https://openthoughts.ai. |
| title | OpenThoughts: Data Recipes for Reasoning Models |
| topic | Machine Learning |
| url | https://arxiv.org/abs/2506.04178 |