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 learning (RL) to improve reasoning capability. DeepSeek-R1 attains results on par with OpenAI's o1 design on a number of criteria, including MATH-500 and SWE-bench.

DeepSeek-R1 is based upon DeepSeek-V3, a mixture of experts (MoE) model just recently open-sourced by DeepSeek. This base model is fine-tuned utilizing Group Relative Policy Optimization (GRPO), a reasoning-oriented variation of RL. The research team likewise carried out understanding distillation from DeepSeek-R1 to open-source Qwen and Llama designs and released numerous variations of each; these designs surpass bigger designs, consisting of GPT-4, on math and coding benchmarks.

[DeepSeek-R1 is] the initial step towards improving language model thinking abilities using pure support knowing (RL). Our objective is to check out the capacity of LLMs to establish reasoning capabilities without any monitored data, focusing on their self-evolution through a pure RL process...DeepSeek-R1 ... master a wide variety of tasks, including imaginative writing, general concern answering, editing, wiki.myamens.com summarization, and more. Additionally, DeepSeek-R1 shows exceptional efficiency on jobs requiring long-context understanding, considerably outshining DeepSeek-V3 on long-context benchmarks.

To develop the design, DeepSeek began with DeepSeek-V3 as a base. They initially tried fine-tuning it just with RL, and with no supervised fine-tuning (SFT), producing a design called DeepSeek-R1-Zero, which they have actually also launched. This design shows strong reasoning efficiency, but" powerful thinking behaviors, it faces a number of issues. For example, DeepSeek-R1-Zero has problem with obstacles like poor readability and language blending."

To address this, the group utilized a short phase of SFT to prevent the "cold start" problem 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 process assembled, they then collected more SFT information using rejection sampling, leading to a dataset of 800k samples. This dataset was used for additional fine-tuning and to produce the distilled models from Llama and Qwen.

DeepSeek evaluated their design on a range of reasoning, math, and coding benchmarks and compared it to other models, consisting of Claude-3.5- Sonnet, GPT-4o, and o1. DeepSeek-R1 exceeded all of them on several of the standards, including AIME 2024 and MATH-500.

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

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

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

Each response starts with a ... pseudo-XML tag containing the chain of thought utilized to help create the response. [Given the prompt] "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 process of getting there was such an interesting insight into how these brand-new designs work.

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

DeepSeek is quickly emerging as a strong builder of open designs. Not only are these models great entertainers, but their license permits usage of their outputs for distillation, potentially pushing forward the state of the art for language designs (and multimodal models) of all sizes.

The DeepSeek-R1 models are available on HuggingFace.