GitLab

DeepSeek open-sourced DeepSeek-R1, an LLM fine-tuned with reinforcement learning (RL) to enhance thinking capability. DeepSeek-R1 attains results on par with OpenAI's o1 model on numerous criteria, including MATH-500 and SWE-bench.

DeepSeek-R1 is based on DeepSeek-V3, a mix of experts (MoE) model just recently open-sourced by DeepSeek. This base design is fine-tuned using Group Relative Policy Optimization (GRPO), a reasoning-oriented variation of RL. The research team likewise performed knowledge distillation from DeepSeek-R1 to open-source Qwen and Llama designs and released several variations of each; these designs exceed bigger models, consisting of GPT-4, on math and coding standards.

[DeepSeek-R1 is] the very first step towards improving language design reasoning capabilities utilizing pure reinforcement learning (RL). Our goal is to check out the potential of LLMs to develop reasoning abilities with no supervised information, focusing on their self-evolution through a pure RL process...DeepSeek-R1 ... excels in a large range of tasks, including imaginative writing, general question answering, editing, summarization, and more. Additionally, DeepSeek-R1 demonstrates impressive efficiency on jobs requiring long-context understanding, considerably surpassing DeepSeek-V3 on long-context benchmarks.

To establish the design, DeepSeek began with DeepSeek-V3 as a base. They initially tried fine-tuning it only with RL, and without any supervised fine-tuning (SFT), a design called DeepSeek-R1-Zero, which they have actually likewise launched. This design shows strong thinking performance, however" effective thinking behaviors, it deals with a number of issues. For example, DeepSeek-R1-Zero deals with difficulties like bad readability and language mixing."

To address this, the team used a short stage of SFT to prevent the "cold start" issue of RL. They collected a number of thousand examples of chain-of-thought thinking to utilize in SFT of DeepSeek-V3 before running RL. After the RL procedure converged, they then collected more SFT information utilizing rejection tasting, resulting in a dataset of 800k samples. This dataset was utilized for additional fine-tuning and to produce the distilled designs from Llama and Qwen.

DeepSeek assessed their design on a range of thinking, mathematics, and coding benchmarks and compared it to other models, including Claude-3.5- Sonnet, GPT-4o, and o1. DeepSeek-R1 surpassed all of them on numerous of the criteria, including AIME 2024 and MATH-500.

DeepSeek-R1 Performance. Image Source: DeepSeek-R1 Technical Report

Within a few days of its release, the LMArena announced that DeepSeek-R1 was ranked # 3 total in the arena and # 1 in coding and mathematics. It was likewise tied for wiki.whenparked.com # 1 with o1 in "Hard Prompt with Style Control" category.

Django framework co-creator Simon Willison wrote about his explores one of the DeepSeek distilled Llama designs on his blog site:

Each action starts with a ... pseudo-XML tag containing the chain of thought utilized to help produce the action. [Given the timely] "a joke about a pelican and a walrus who run a tea room together" ... It then thought for 20 paragraphs before outputting the joke! ... [T] he joke is terrible. But the procedure of arriving was such an interesting insight into how these new designs work.

Andrew Ng's newsletter The Batch discussed DeepSeek-R1:

DeepSeek is quickly becoming a strong contractor of open models. Not just are these models great entertainers, however their license allows usage of their outputs for distillation, possibly pushing forward the cutting-edge for language models (and multimodal models) of all sizes.

The DeepSeek-R1 designs are available on HuggingFace.

About the Author

Anthony Alford

Rate this Article

This material remains in the AI, ML & Data Engineering subject

Related Topics:

- AI, ML & Data Engineering - Generative AI

• Large language models
  
  - Related Editorial
  
  Related Sponsored Content
  
  - [eBook] Getting Going with Azure Kubernetes Service
  
  Related Sponsor
  
  Free services for AI apps. Are you all set to experiment with innovative technologies? You can begin developing smart apps with totally free Azure app, information, and AI services to minimize upfront expenses. Find out more.
  
  How could we enhance? Take the InfoQ reader survey
  
  Each year, we look for feedback from our readers to help us improve InfoQ. Would you mind costs 2 minutes to share your feedback in our brief survey? Your feedback will straight help us continuously progress how we support you. The InfoQ Team Take the study
  
  Related Content
  
  The InfoQ Newsletter
  
  A round-up of last week's material on InfoQ sent every Tuesday. Join a neighborhood of over 250,000 senior developers.