What is Sky-T1
Sky-T1 is an open-source reasoning AI model released by NovaSky, a research team from the Sky Computing Lab at UC Berkeley, named Sky-T1-32B-Preview. It is the first open-source reasoning model, with both the training dataset and code publicly available, allowing users to reproduce the model from scratch. The model's training cost is less than $450. Sky-T1's training data was generated by Alibaba'sQwQ-32B-Previewreasoning model, carefully filtered and reconstructed, and processed based on OpenAI's GPT-4o-mini, making it easier for model training. In terms of performance, Sky-T1 outperforms the preview version of OpenAI's earlier o1 on MATH500 (a set of "competition-level" math challenges), and also performs better in programming evaluations on LiveCodeBench.
Main features of Sky-T1
- Open source: Sky-T1's training dataset and code are both publicly available, allowing users to reproduce the model from scratch.
- Low-cost training: The model's training cost is less than $450, significantly lower than the millions of dollars in training costs of previous models with comparable performance.
- Reasoning capability: Sky-T1 can effectively conduct self-fact-checking, avoid some common pitfalls, and may take longer to arrive at solutions, ranging from a few seconds to several minutes. Its reliability in fields such as physics, science, and mathematics is higher.
- Training data and hardware: Sky-T1's training data was generated by Alibaba's QwQ-32B-Preview reasoning model, carefully filtered, and reconstructed based on OpenAI's GPT-4o-mini, making it easier to process. Training this model with 32 billion parameters took only about 19 hours, using 8 Nvidia H100 GPUs.
- Performance: It outperforms the early preview version of OpenAI's o1 on MATH500 (a set of "competition-level" math challenges), and also beats o1's preview version on a set of difficult problems from LiveCodeBench (a coding evaluation).
Sky-T1 project addresses
- Project official website：https://novasky-ai.github.io/posts/sky-t1
- HuggingFace model repository：https://huggingface.co/NovaSky-AI/Sky-T1-32B-Preview
Application scenarios of Sky-T1
- Mathematical problem solving: Sky-T1 outperforms the preview version of OpenAI's earlier o1 on MATH500 (a set of "competition-level" math challenges), has significant advantages in solving complex mathematical problems, and is suitable for fields such as education, scientific research, and competitions.
- Programming evaluation: In LiveCodeBench's programming evaluation, Sky-T1 also performs excellently, can effectively help developers with code generation, optimization, and error detection, and is suitable for scenarios such as software development, programming education, and automated testing.
- Scientific research: Although Sky-T1 is slightly inferior to the o1 preview version in the GPQA-Diamond test (which includes graduate-level questions in the fields of physics, biology, and chemistry), its reliability in fields such as physics, science, and mathematics is higher, making it suitable for solving complex problems in scientific research and academic studies.