open-thoughts/OpenThinker3-1.5B
16134
1---2base_model: Qwen/Qwen2.5-1.5B-Instruct3datasets:4- open-thoughts/OpenThoughts3-1.2M5library_name: transformers6license: apache-2.07tags:8- llama-factory9- full10- generated_from_trainer11model-index:12- name: OpenThinker3-7B13 results: []14pipeline_tag: text-generation15---16 17<p align="center">18 <img src="https://huggingface.co/datasets/open-thoughts/open-thoughts-114k/resolve/main/open_thoughts.png" width="50%">19</p>20 21<p align="center">22<a href="https://arxiv.org/abs/2506.04178" style="margin-right: 24px;">paper</a> |23<a href="https://huggingface.co/datasets/open-thoughts/OpenThoughts3-1.2M" style="margin-right: 24px; margin-left: 24px;">dataset</a> |24<a href="https://huggingface.co/open-thoughts/OpenThinker3-7B" style="margin-left: 24px;">model</a>25</p>26 27> [!NOTE]28> We have released a paper for OpenThoughts! See our paper [here](https://arxiv.org/abs/2506.04178).29 30# OpenThinker3-1.5B31 32State-of-the-art SFT-only 1.5B reasoning model. 🚀33 34This model is a fine-tuned version of [Qwen/Qwen2.5-1.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct) on the 35[OpenThoughts3-1.2M](https://huggingface.co/datasets/open-thoughts/OpenThoughts3-1.2M) dataset. 36 37See our [paper](https://arxiv.org/abs/2506.04178) and [blog post](https://openthoughts.ai/blog/ot3) for more details.38 39# Evaluation Results40The numbers reported in the table below are evaluated with our open-source tool [Evalchemy](https://github.com/mlfoundations/Evalchemy). 41In the table below, we bold values in each column that are within 2 standard errors of the best.42 43| Model | AIME24 | AIME25 | AMC23 | MATH500 | HMMT O2/25 | LCB 06/24-01/25 | CodeElo | CodeForces | GPQA-D | JEEBench |44| ------------------------------------------------------------------------------------------------------------ | ------ | ------ | ------ | ------- | ---------- | --------------- | ------- | ---------- | ------ | -------- |45| **[OpenThinker3-1.5B](https://huggingface.co/open-thoughts/OpenThinker3-1.5B)** |**52.0**|**41.7**|**87.0**| 86.4 | **27.3** | **39.4** | 12.9 | 15.5 | 29.5 | 51.9 |46| [DeepSeek-R1-Distill-Qwen-1.5B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B) | 32.3 | 23.7 | 71.8 | 80.8 | 15.3 | 27.2 | 8.8 | 8.5 | 31.1 | 32.5 |47| [Nemotron-Research-Reasoning-Qwen-1.5B](https://huggingface.co/nvidia/Nemotron-Research-Reasoning-Qwen-1.5B) |**47.7**| 32.0 |**87.5**| 86.0 | 21.7 | 31.4 |**54.7** |**40.3** | 41.8 | 52.6 |48| [Qwen3-1.7B](https://huggingface.co/Qwen/Qwen3-1.7B) |**52.0**| 35.3 | 83.8 | **87.2**| 23.3 | 27.7 | 20.7 | 20.0 |**49.3**|**60.7** |49| [Qwen2.5-1.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct) | 3.0 | 0.7 | 30.8 | 50.2 | 0.0 | 5.5 | 0.8 | 2.2 | 24.7 | 16.4 |50 51# Data52 53This model was trained on the [OpenThoughts3-1.2M](https://huggingface.co/datasets/open-thoughts/OpenThoughts3-1.2M) dataset.54 55The key to the strong model performance is our comprehensive data pipeline and over 1,000+ ablation experiments.56This led to the creation of [OpenThoughts3-1.2M](https://huggingface.co/datasets/open-thoughts/OpenThoughts3-1.2M), which consists of 850,000 math questions, 250,000 code questions, and 100,000 science questions.57Reasoning traces are generated with QwQ-32B. 58 59See the [OpenThoughts3-1.2M](https://huggingface.co/datasets/open-thoughts/OpenThoughts3-1.2M) dataset page or our [paper](https://arxiv.org/abs/2506.04178) for additional information.60 61 62# Intended uses & limitations63 64Apache 2.0 License65 66## Training procedure67 68We used 64 A100 gpus to train the model for 7 days.69 70## Training hyperparameters71 72The following hyperparameters were used during training:73- learning_rate: 0.0001674- train_batch_size: 475- eval_batch_size: 876- seed: 4277- distributed_type: multi-GPU78- num_devices: 6479- total_train_batch_size: 25680- optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments81- lr_scheduler_type: cosine82- lr_scheduler_warmup_ratio: 0.183- num_epochs: 7.084 85## Framework versions86 87- Transformers 4.46.188- Pytorch 2.3.089- Datasets 3.1.090- Tokenizers 0.20.391 92More info can be found in our repository: [https://github.com/open-thoughts/open-thoughts](https://github.com/open-thoughts/open-thoughts).93 94# Links95- 📝 [OpenThoughts Paper](https://arxiv.org/abs/2506.04178)96- 📊 [OpenThoughts3-1.2M and OpenThinker3-7B Blog Post](https://www.open-thoughts.ai/blog/ot3)97- 💻 [Open Thoughts GitHub Repository](https://github.com/open-thoughts/open-thoughts)98- 🧠 [OpenThoughts3-1.2M dataset](https://huggingface.co/datasets/open-thoughts/OpenThoughts3-1.2M)99- 🤖 [OpenThinker3-7B model](https://huggingface.co/open-thoughts/OpenThinker3-7B)100- 🤖 [OpenThinker3-1.5B model](https://huggingface.co/open-thoughts/OpenThinker3-1.5B) - this model.101 102 103# Citation104```105@misc{guha2025openthoughtsdatarecipesreasoning,106 title={OpenThoughts: Data Recipes for Reasoning Models}, 107 author={Etash Guha and Ryan Marten and Sedrick Keh and Negin Raoof and Georgios Smyrnis and Hritik Bansal and Marianna Nezhurina and Jean Mercat and Trung Vu and Zayne Sprague and Ashima Suvarna and Benjamin Feuer and Liangyu Chen and Zaid Khan and Eric Frankel and Sachin Grover and Caroline Choi and Niklas Muennighoff and Shiye Su and Wanjia Zhao and John Yang and Shreyas Pimpalgaonkar and Kartik Sharma and Charlie Cheng-Jie Ji and Yichuan Deng and Sarah Pratt and Vivek Ramanujan and Jon Saad-Falcon and Jeffrey Li and Achal Dave and Alon Albalak and Kushal Arora and Blake Wulfe and Chinmay Hegde and Greg Durrett and Sewoong Oh and Mohit Bansal and Saadia Gabriel and Aditya Grover and Kai-Wei Chang and Vaishaal Shankar and Aaron Gokaslan and Mike A. Merrill and Tatsunori Hashimoto and Yejin Choi and Jenia Jitsev and Reinhard Heckel and Maheswaran Sathiamoorthy and Alexandros G. Dimakis and Ludwig Schmidt},108 year={2025},109 eprint={2506.04178},110 archivePrefix={arXiv},111 primaryClass={cs.LG},112 url={https://arxiv.org/abs/2506.04178}, 113}114```