DeepSeek Math 7B Instruct

Overview

DeepSeek Math 7B Instruct is the instruction-tuned member of DeepSeek's DeepSeekMath family. Its base, DeepSeekMath-Base 7B, started from DeepSeek-Coder-Base-v1.5 7B and was trained on another 500B tokens, 56% of them from a 120B-token math corpus (35.5M web pages) that DeepSeek collected from Common Crawl.

Instruction tuning then used 776K examples built from English and Chinese math problems, with solutions written as chain-of-thought, program-of-thought and tool-integrated reasoning. DeepSeek reports 46.8% on MATH and 82.9% on GSM8K for this model with chain-of-thought prompting and no tools.

What you can build

• Tutoring tools that show every line of a worked solution, so a student sees where an equation was rearranged.
• Answer checkers that read the final answer out of the \boxed{} wrapper and compare it with a key.
• Math help for students who ask in Chinese as well as English.
• Draft solutions for a problem set, reviewed by a teacher before they go out.
• Follow-up questions in one session, where the model still has the earlier exchange.

Inputs

The only required field is the question, as plain text. LaTeX is fine: the default prompt asks for the x that satisfies 4x + 5 = 6x + 7, written in LaTeX math.

Advanced settings include a system prompt (prefilled with a generic assistant instruction), temperature from 0 to 5 (default 0.7), top-p, top-k, repetition and length penalties, token limits, semicolon-separated stop sequences and a seed (default 123456). Sampling and quantization are switches that start on. Optional user_id and session_id values keep a separate chat history per user and per conversation.

Outputs

You get text back: the solution worked through one step at a time. When the question ends with DeepSeek's recommended instruction, the final answer sits inside \boxed{}, which a script can find with a simple pattern match.

Recommended settings

• End English questions with "Please reason step by step, and put your final answer within \boxed{}." For Chinese questions DeepSeek gives "请通过逐步推理来解答问题，并把最终答案放置于\boxed{}中。"
• Empty the system prompt field. DeepSeek says this version is not compatible with a system prompt and advises against using one.
• Switch sampling off when answers are graded or must come out the same on every run.
• Quantization is on by default, and its own note says it can slightly affect output quality. Switch it off to run the model at full precision.

Limitations

• DeepSeekMath models have a 4,096-token sequence length, so a long problem set or chat history may not fit.
• Part of the tuning data taught the model to answer by writing a program for a tool to run. Nothing executes code on this page: any program comes back as text for you to run.
• DeepSeek's 46.8% on MATH means more than half of those competition problems came back wrong. Check final answers before you present them as correct.
• The instruction data and DeepSeek's suggested prompts are in English and Chinese.

Safety & compliance

Use of the weights is governed by the DeepSeek License Agreement 1.0, and DeepSeek states that DeepSeekMath supports commercial use. Attachment A of that license bars military use and fully automated decision making that adversely affects a person's legal rights, along with other listed restrictions.