O1-Journey is an open research project centered on reproducing the OpenAI O1 model, advanced by the core development team of the GAIR research group at Shanghai Jiao Tong University. Team members are mainly third- and fourth-year undergraduates and first-year doctoral students, and are guided by research scientists in the large language model field from institutions such as New York University and MBZUAI.
In terms of purpose, the project aims to reproduce O1's capabilities through a transparent, real-time exploratory approach, while rethinking how AI research is conducted and communicated. The project proposes "journey learning," a new training paradigm that emphasizes improving an AI system's ability to handle complex real-world problems through continuous learning, reflection, and adaptation, rather than focusing only on specific tasks. This paradigm combines search and learning; the search process includes trial and error, correction, backtracking, and reflection, and is considered an effective path for complex reasoning tasks.
The main capabilities and results are presented in three parts. Part1 is a strategic progress report, proposing journey learning and releasing the journey thought training dataset, while also providing the dataset, Walnut Plan, and paper citation information. Part2 studies reproducing O1 through simple distillation, pointing out that distilling from the O1 API combined with supervised fine-tuning can achieve performance superior to O1-preview on complex mathematical reasoning tasks; key findings include: a base model fine-tuned on tens of thousands of O1 distillation chains surpasses O1-preview on AIME with extremely low technical complexity; although trained only on mathematical data, the distilled model performs strongly in open-domain question answering, reducing hallucinations, improving safety, and lowering sensitivity to sycophantic behavior; the paper also emphasizes transparent research practices and prioritizing first-principles thinking. Part3 explores the application of inference-time scaling in medical reasoning, based on previous journey learning and distillation work, studying the impact of extending inference time on medical task performance; experiments on benchmarks such as MedQA, Medbullets, and JAMA Clinical Challenges show that using only 500 training samples can bring a 6%-11% performance improvement; the model's differential diagnosis process is similar to the hypothetico-deductive method used by medical professionals, systematically evaluating evidence to narrow the range of possible conditions.
In terms of usage, project resources are gradually released through the GitHub repository, datasets are published on Hugging Face, and related reports and paper citation information are provided in the repository. The project states that it will continue to share new progress, with more resources coming soon.
Applicable scenarios include complex reasoning research on large language models, model training and evaluation for mathematical reasoning and medical reasoning, as well as transparent exploration of AI research methods.
In terms of limitations, the materials do not provide specific installation commands, code usage steps, fees, or pricing information, nor do they explain platform compatibility or deployment requirements. The project is research-oriented, and actual use requires judgment based on the reports, datasets, and papers in the repository.
