Sky-T1-32B-Flash Open-Source Inference Model - The Optimized Version Significantly Shortens the Generation Length while Maintaining Accuracy

Sky T1 32B Flash

Developed by NovaSky-AI

An optimized 32B inference model preference version based on Sky-T1-32B-Preview, significantly reducing generation length while maintaining accuracy.

Large Language Model

Transformers

EnglishOpen Source License:Apache-2.0 #Mathematical Reasoning Optimization #Programming Task Efficiency #Generation Length Compression

Downloads 557

Release Time : 1/23/2025

Model Overview

This model performs comparably to the o1-preview model in mathematical and programming tasks while reducing generation length by up to 57% compared to Sky-T1-32B-Preview.

Model Features

Efficient Inference

Reduces generation length by up to 57% compared to Sky-T1-32B-Preview, significantly lowering inference costs.

Accuracy Retention

Performs comparably to the o1-preview model in mathematical and programming tasks without compromising accuracy.

Optimized Training

Trained using the Simple Preference Optimization (SimPO) method to enhance model efficiency.

Model Capabilities

Mathematical reasoning

Programming task reasoning

Text generation

Use Cases

Mathematics

Mathematical Problem Solving

Solves complex mathematical problems, such as AIME24 competition questions.

Accuracy rate 43.3%, average length reduced by 37%.

Programming

Programming Problem Solving

Solves easy, medium, and hard problems in programming competitions.

Easy problems accuracy 89%, medium 56.3%, hard 17.9%, average length reduced by 34%-57%.

Knowledge QA

MMLU Knowledge QA

Answers multidisciplinary knowledge questions.

Accuracy rate 81.7%, average length reduced by 17%.

🚀 Sky-T1-32B-Flash Model

This is a 32B reasoning model that optimizes preferences on top of Sky-T1-32B-Preview. It significantly reduces generation lengths while maintaining accuracy, offering high - performance reasoning capabilities.

🚀 Quick Start

The following sections will introduce the details of the Sky-T1-32B-Flash model, including model details, training details, evaluation results, etc.

✨ Features

Optimized on Sky-T1-32B-Preview to reduce generation lengths by up to 57% while maintaining accuracy.
Performs on par with the o1 - preview model in math and coding.
Uses Simple Policy Optimization (SimPO) for training.

📚 Documentation

📋 Model Details

Model Description

This is a 32B reasoning model preference optimized on top of Sky-T1-32B-Preview to significantly reduce generation lengths while maintaining accuracy. The performance is on par with o1-preview model in both math and coding, while reducing generation lengths by up to 57% relative to Sky-T1-32B-Preview. Please see our blog post for more details.

Developed by: NovaSky Team from Sky Computing Lab at UC Berkeley.

📈 Training Details

Training Data

10K preference pairs in math and coding domains, generated by Sky-T1-32B-Preview.

Training Procedure

We perform Simple Policy Optimization (SimPO) with a batch size of 96, learning rate of 5e-7, gamma of 0.3, and beta of 2.0.

Speeds

We use Llama-Factory for training. On 8xH100, the SimPO training takes ~2.5 hours with DeepSpeed Zero-3 Offload.

🧪 Evaluation

Property	Sky-T1-32B-Preview	Sky-T1-32B-Flash	Qwen2.5-32B-Instruct	QwQ-32B-Base	DeepSeek-R1-Distill-Qwen-32B
Math500 Acc	88.6	88.6	76.2	89.2	90.8
Math500 Avg Len	2124	1417 (-33%)	522	2089	2010
AIME24 Acc	43.3	43.3	16.7	50	66.7
AIME24 Avg Len	6881	4365 (-37%)	970	7379	9173
LCB Easy Acc	87.4	89	84.6	90.7	91.2
LCB Easy Avg Len	3415	2265 (-34%)	414	3255	2775
LCB Medium Acc	56.8	56.3	40.8	56.3	76.7
LCB Medium Avg Len	8263	4389 (-47%)	535	6742	6324
LCB Hard Acc	17.9	17.9	9.8	17.1	38.2
LCB Hard Avg Len	14564	6199 (-57%)	618	10450	10448
MMLU Acc	82.4	81.7	80.1	85.2	82.1
MMLU Avg Len	1087	799 (-17%)	312	1041	774
GPQA Diamond Acc	56.8	56.6	45.5	52.5	62.6
GPQA Diamond Avg Len	3503	2148 (-39%)	600	3302	5108

🙏 Acknowledgement

We would like to thanks the compute resources from Lambda Lab and AnyScale.

📄 License

Apache-2.0

📖 Citation

Please considering citing our blog post if you found it useful for your research. Thank you!

@misc{reduce_overthinking_2025,
  author       = {NovaSky Team},
  title        = {Think Less, Achieve More: Cut Reasoning Costs by 50% Without Sacrificing Accuracy},
  howpublished = {https://novasky-ai.github.io/posts/reduce-overthinking},
  note         = {Accessed: 2025-01-23},
  year         = {2025}
}

Featured Recommended AI Models

Empowering the Future, Your AI Solution Knowledge Base

English 简体中文繁體中文にほんご