DeepSeek Open-Sources DeepSeek-R1 LLM with Performance Comparable To OpenAI's O1 Model

DeepSeek open-sourced DeepSeek-R1, an LLM fine-tuned with reinforcement knowing (RL) to improve thinking capability. DeepSeek-R1 attains results on par with OpenAI's o1 design on numerous standards, consisting of MATH-500 and SWE-bench.

DeepSeek-R1 is based upon DeepSeek-V3, a mix of experts (MoE) design recently open-sourced by DeepSeek. This base design is fine-tuned using Group Relative Policy Optimization (GRPO), a reasoning-oriented version of RL. The research group also carried out understanding distillation from DeepSeek-R1 to open-source Qwen and Llama models and released a number of versions of each; these models outshine bigger models, larsaluarna.se consisting of GPT-4, on mathematics and coding benchmarks.

[DeepSeek-R1 is] the primary step toward enhancing language model thinking abilities utilizing pure support learning (RL). Our objective is to explore the potential of LLMs to establish thinking abilities with no supervised data, focusing on their self-evolution through a pure RL process...DeepSeek-R1 ... excels in a wide variety of tasks, including creative writing, general question answering, modifying, summarization, and more. Additionally, DeepSeek-R1 demonstrates exceptional efficiency on tasks needing long-context understanding, considerably surpassing DeepSeek-V3 on long-context criteria.

To develop the design, DeepSeek started with DeepSeek-V3 as a base. They first tried fine-tuning it just with RL, and without any supervised fine-tuning (SFT), producing a model called DeepSeek-R1-Zero, which they have also launched. This model displays strong thinking performance, however" powerful reasoning habits, it deals with a number of problems. For instance, DeepSeek-R1-Zero battles with challenges like poor readability and language mixing."

To address this, the team utilized a short phase of SFT to avoid the "cold start" issue of RL. They gathered several thousand examples of chain-of-thought thinking to utilize in SFT of DeepSeek-V3 before running RL. After the RL process converged, they then gathered more SFT information utilizing rejection sampling, leading to a dataset of 800k samples. This dataset was utilized for additional fine-tuning and to produce the distilled models from Llama and Qwen.

DeepSeek examined their design on a variety of thinking, math, and coding standards 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, consisting of 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 overall in the arena and # 1 in coding and math. It was likewise tied for # 1 with o1 in "Hard Prompt with Style Control" category.

Django framework co-creator Simon Willison wrote about his experiments with one of the DeepSeek distilled Llama models on his blog:

Each action starts with a ... pseudo-XML tag containing the chain of idea used to assist produce the response. [Given the prompt] "a joke about a pelican and a walrus who run a tea space together" ... It then believed for 20 paragraphs before outputting the joke! ... [T] he joke is dreadful. But the procedure of arriving was such an intriguing insight into how these brand-new designs work.

Andrew Ng's newsletter The Batch blogged about DeepSeek-R1:

DeepSeek is rapidly emerging as a strong builder of open designs. Not just are these designs fantastic entertainers, however their license permits use of their outputs for distillation, possibly pressing forward the state of the art for language models (and multimodal designs) of all sizes.

The DeepSeek-R1 designs are available on HuggingFace.