Sky-T1 is NovaSky's open-source inference AI model, which can be reproduced from scratch.
Sky-T1 is an open-source inference AI model released by NovaSky, a research team at the Sky Computing Lab at UC Berkeley. It is named Sky-T1-32B-Preview. It is the first open-source inference model, with both its training dataset and code...
What is Sky-T1?
Sky-T1 is an open-source inference AI model released by NovaSky, a research team at the Sky Computing Lab at UC Berkeley. Originally named Sky-T1-32B-Preview, it is the first open-source inference model with its training dataset and code publicly available, allowing users to reproduce the model from scratch. The training cost is less than $450. Sky-T1's training data was generated from Alibaba's QwQ-32B-Preview inference model, carefully selected and reconstructed, and processed using OpenAI's GPT-4o-mini, making it easier to train. In terms of performance, Sky-T1 outperforms the preview version of OpenAI's earlier version o1 on MATH500 (a set of "competition-level" mathematical challenges) and also performs better in LiveCodeBench programming evaluations.
Key features of Sky-T1
- open sourceThe training dataset and code for Sky-T1 are publicly available, allowing users to reproduce the model from scratch.
- Low-cost trainingThe training cost of the model is less than $450, significantly lower than the millions of dollars required to train a comparable model in the past.
- reasoning abilitySky-T1 can effectively perform self-fact checking, avoiding some common pitfalls, although it may take longer, ranging from seconds to minutes, to arrive at a solution. It is more reliable in fields such as physics, science, and mathematics.
- Training data and hardwareThe training data for Sky-T1 was generated from Alibaba's QwQ-32B-Preview inference model. After careful selection, the data was 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 and used 8 Nvidia H100 GPUs.
- PerformanceIt outperformed an early preview version of OpenAI's o1 on MATH500 (a set of "competition-level" math challenges) and also beat the o1 preview version on a set of problems from LiveCodeBench (a coding evaluation tool).
Sky-T1 project address
- Project official website:https://novasky-ai.github.io/posts/sky-t1
- HuggingFace model library:https://huggingface.co/NovaSky-AI/Sky-T1-32B-Preview
Application scenarios of Sky-T1
- Mathematical Problem SolvingSky-T1 outperforms the preview version of earlier OpenAI version o1 on MATH500 (a set of “competition-level” mathematical challenges), demonstrating a significant advantage in solving complex mathematical problems and making it suitable for fields such as education, research, and competitions.
- Programming AssessmentIn LiveCodeBench's programming assessment, Sky-T1 also performed excellently, effectively helping developers with code generation, optimization, and error detection, making it suitable for software development, programming education, and automated testing scenarios.
- Scientific researchAlthough Sky-T1 is slightly inferior to the o1 preview version in the GPQA-Diamond test (which includes graduate-level problems in physics, biology and chemistry), it is more reliable in physics, science and mathematics and is suitable for solving complex problems in scientific and academic research.