Introduction to DeepSeek-Math
DeepSeekMath is an advanced machine learning model aimed at improving mathematical problem-solving capabilities by leveraging a large-scale pre-trained architecture. It builds upon the foundation of DeepSeek-Coder-v1.5 7B and specializes in comprehensively handling math-related tasks by incorporating a vast amount of math-related data. The model achieves a remarkable performance score of 51.7% on high-level math benchmarks, approaching the capabilities of leading models like Gemini-Ultra and GPT-4, without the need for additional tools or voting techniques. Researchers have access to a variety of model versions, including base, instruct, and RL models.
Evaluation Results
DeepSeekMath-Base 7B
The base version of DeepSeekMath is meticulously tested for its mathematical abilities, focusing on:
- Solving mathematical problems independently
- Utilizing external tools for complex calculations
- Engaging in formal theorem proving
In addition to math, it showcases proficiencies in:
- Language comprehension and logic
- Programming and coding capabilities
Key performance highlights:
- Mathematical Reasoning: It delivers outstanding results on competitive math datasets, performing over 10% better than other open-source models using few-shot prompting.
- Tool Usage: Its training regimen enhances its ability to write programs that solve and verify math problems.
- Reasoning and Coding: It offers reasoning and coding performance on par with the original DeepSeekCoder-Base version.
DeepSeekMath-Instruct and -RL 7B
DeepSeekMath-Instruct 7B is derived from the base model, emphasizing mathematical instruction, while DeepSeekMath-RL 7B integrates an advanced training algorithm, GRPO. The models are evaluated across various benchmarks in both English and Chinese contexts, exhibiting strong performance in step-by-step reasoning and surpassing other models in tool-assisted tasks.
Data Collection Process
The data collection for DeepSeekMath involved a meticulous, five-step process:
- Seed Corpus Selection: Utilized OpenWebMath as the basis for training a FastText model.
- Web Page Retrieval: Extracted mathematical content from Common Crawl.
- Domain Identification: Analyzed data to determine math-related domains.
- URL Annotation: Marked URLs with potential math content.
- Content Enrichment: Integrated additional web pages iteratively to expand the dataset.
This process resulted in a comprehensive collection of 35.5 million web pages, translating to a significant data volume of 120 billion tokens.
Model Downloads
Public access to DeepSeekMath's various models (base, instruct, RL) through platforms like Huggingface ensures widespread application and research potential in both academic and commercial settings, under specific licensing terms.
Quick Start
For those interested in employing the DeepSeekMath models, Huggingface's Transformers library provides a straightforward method for model inference. Users can engage with the models through text completion and chat completion scripts, allowing for practical exploration of its capabilities.
License
The model's code is available under the MIT License, while the models themselves are governed by a dedicated Model License, permitting commercial usage under defined conditions.
Citation
For acknowledgment in academic usage, a proper citation format is provided to reference the work on DeepSeekMath.
Contact
Inquires and issues can be directed to the DeepSeekMath team via email, strengthening collaboration and support for users around the globe.