DeepSeek Official Update on Skipping Reasoning Before Output 191

小编 ·

DeepSeek Official Update on Skipping Reasoning Before Output

DeepSeek has updated its usage guide for R1. Previously, users of our fully deployed deepseek-r1 cluster frequently encountered situations where the model output reasoning results directly without any thinking process. DeepSeek's official solution is as follows:

DeepSeek Official Update on Skipping Reasoning Before Output

DeepSeek has officially updated the usage instructions for R1. Previously, the full-capacity deepseek-r1 deployed on our cluster would frequently output reasoning results directly without thinking during multi-user sessions. To address this issue, DeepSeek officially provides the following solutions:

1. Please set the Temperature within the range of 0.5 to 0.7 (0.6 is recommended) to avoid endless repetition or incoherent outputs.

2. Do not add system prompts; all instructions should be included in the user prompt.

3. For math problems, it is recommended to add an instruction prompt in your prompt: "Please reason step by step, and put your final answer within \boxed{}."

4. When evaluating model performance, it is recommended to conduct multiple tests and average the results.

5. When responding to certain queries, the DeepSeek-R1 series models tend to bypass the deep thinking mode (i.e., output ""), which may adversely affect model performance. To ensure the model can perform comprehensive reasoning, the official recommendation is: enforce "" at the beginning of each output

The situation where reasoning results are output directly without thinking, DeepSeek official has made the latest usage instructions.webp

6. Officially recommended search and file upload prompt template: https://github.com/deepseek-ai/DeepSeek-R1/pull/399/files

7. DeepSeek-R1 guide on mitigating the model skipping thinking and reasoning directly: https://github.com/deepseek-ai/DeepSeek-R1/commit/7ca5e1e7f75e12a1c561fffaa6aa686708f881ae

DeepSeek official has also updated the readme. You can follow the official readme to configure and get the best user experience.

Latest readme address: https://github.com/deepseek-ai/DeepSeek-R1/blob/main/README.md