DeepSeek-V3-0324-Fused-4E-29B-Unhealed-Preview

<!-- markdownlint-disable first-line-h1 --> <!-- markdownlint-disable html --> <!-- markdownlint-disable no-duplicate-header -->

README.md

Unhealed DeepSeek-v3-0324-Instruct Fused Models (Research Release)

CRITICAL NOTE: Untrained Fusion - Requires Healing!

These are unhealed, experimental versions of DeepSeek-v3-0324-instruct created through model fusion. They are not ready for direct use and will exhibit unpredictable behavior without significant post-training. These models are released exclusively for research purposes and require a specific "healing" process to restore functionality. Do not use these models without understanding and applying the healing procedure.

Preview Models: Exploring Compression Levels

The DeepSeek-V3-0324 model, which utilizes 256 experts, forms the foundation for these preview models. We offer four variations, each with a different level of compression:

• 8 Fused Experts, rank 4 (~39B parameters): 1/20 size reduction.
• 4 Fused Experts, rank 4 (~29B parameters): 1/23 size reduction.

Despite their significantly reduced size, these models demonstrate surprisingly strong performance, exceeding expectations for their parameter counts. Further, more comprehensive testing is planned.

What to Expect (Before Healing)

These models are in an initial, unstable state after the fusion process. Expect significantly degraded performance and unpredictable outputs. They are not representative of the final capabilities of a properly trained fused model. This is a very early iteration of the fusion and distillation technique, using a small sample size for distillation. Significant room for improvement remains in the distillation process.

Healing Instructions (Required)

Crucially, you must perform post-training to make these models usable. The necessary scripts and detailed instructions are available in the moe-pruner repository:

https://github.com/gabrielolympie/moe-pruner

Follow the instructions in that repository carefully to "heal" the pruned model. This process is essential to recover performance.

Contributing and Future Improvements

This release represents an initial exploration of model fusion and distillation. Due to hardware limitations, significant compromises were made during development.

We welcome contributions to improve this work! There are two primary ways to help:

1. Financial Support: Larger-scale experiments require significant compute resources. If you'd like to support future versions with a higher compute budget, you can donate here: https://gofund.me/1516dccd

2. Code Contributions: Suggest improvements, bug fixes, or new features directly on the GitHub repository.

We are actively working to improve the fusion and distillation techniques, and your contributions are greatly appreciated.

Disclaimer

These models are provided "as is" for research purposes only. No guarantees are made regarding their performance or stability before the healing process is completed. Use at your own risk.

Original Model Card

---

license: mit library_name: transformers

DeepSeek-V3-0324

<!-- markdownlint-disable first-line-h1 --> <!-- markdownlint-disable html --> <!-- markdownlint-disable no-duplicate-header --> <div align="center"> <img src="https://github.com/deepseek-ai/DeepSeek-V2/blob/main/figures/logo.svg?raw=true" width="60%" alt="DeepSeek-V3" /> </div> <hr> <div align="center" style="line-height: 1;"> <a href="https://www.deepseek.com/" target="_blank" style="margin: 2px;"> <img alt="Homepage" src="https://github.com/deepseek-ai/DeepSeek-V2/blob/main/figures/badge.svg?raw=true" style="display: inline-block; vertical-align: middle;"/> </a> <a href="https://chat.deepseek.com/" target="_blank" style="margin: 2px;"> <img alt="Chat" src="https://img.shields.io/badge/🤖%20Chat-DeepSeek%20V3-536af5?color=536af5&logoColor=white" style="display: inline-block; vertical-align: middle;"/> </a> <a href="https://huggingface.co/deepseek-ai" target="_blank" style="margin: 2px;"> <img alt="Hugging Face" src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-DeepSeek%20AI-ffc107?color=ffc107&logoColor=white" style="display: inline-block; vertical-align: middle;"/> </a> </div> <div align="center" style="line-height: 1;"> <a href="https://discord.gg/Tc7c45Zzu5" target="_blank" style="margin: 2px;"> <img alt="Discord" src="https://img.shields.io/badge/Discord-DeepSeek%20AI-7289da?logo=discord&logoColor=white&color=7289da" style="display: inline-block; vertical-align: middle;"/> </a> <a href="https://github.com/deepseek-ai/DeepSeek-V2/blob/main/figures/qr.jpeg?raw=true" target="_blank" style="margin: 2px;"> <img alt="Wechat" src="https://img.shields.io/badge/WeChat-DeepSeek%20AI-brightgreen?logo=wechat&logoColor=white" style="display: inline-block; vertical-align: middle;"/> </a> <a href="https://twitter.com/deepseek_ai" target="_blank" style="margin: 2px;"> <img alt="Twitter Follow" src="https://img.shields.io/badge/Twitter-deepseek_ai-white?logo=x&logoColor=white" style="display: inline-block; vertical-align: middle;"/> </a> </div> <div align="center" style="line-height: 1;"> <a href="LICENSE" style="margin: 2px;"> <img alt="License" src="https://img.shields.io/badge/License-MIT-f5de53?&color=f5de53" style="display: inline-block; vertical-align: middle;"/> </a> </div>

Features

DeepSeek-V3-0324 demonstrates notable improvements over its predecessor, DeepSeek-V3, in several key aspects.

Reasoning Capabilities

• Significant improvements in benchmark performance:
  • MMLU-Pro: 75.9 → 81.2 (+5.3)
  • GPQA: 59.1 → 68.4 (+9.3)
  • AIME: 39.6 → 59.4 (+19.8)
  • LiveCodeBench: 39.2 → 49.2 (+10.0)

Front-End Web Development

• Improved the executability of the code
• More aesthetically pleasing web pages and game front-ends

Chinese Writing Proficiency

• Enhanced style and content quality:

  • Aligned with the R1 writing style
  • Better quality in medium-to-long-form writing

• Feature Enhancements

  • Improved multi-turn interactive rewriting
  • Optimized translation quality and letter writing

Chinese Search Capabilities

• Enhanced report analysis requests with more detailed outputs

Function Calling Improvements

• Increased accuracy in Function Calling, fixing issues from previous V3 versions

---

Usage Recommendations

System Prompt

In the official DeepSeek web/app, we use the same system prompt with a specific date.

该助手为DeepSeek Chat，由深度求索公司创造。
今天是{current date}。

For example,

该助手为DeepSeek Chat，由深度求索公司创造。
今天是3月24日，星期一。

Temperature

In our web and application environments, the temperature parameter $T_{model}$ is set to 0.3. Because many users use the default temperature 1.0 in API call, we have implemented an API temperature $T_{api}$ mapping mechanism that adjusts the input API temperature value of 1.0 to the most suitable model temperature setting of 0.3.

$$ T_{model} = T_{api} \times 0.3 \quad (0 \leq T_{api} \leq 1) $$

$$ T_{model} = T_{api} - 0.7 \quad (1 < T_{api} \leq 2) $$

Thus, if you call V3 via API, temperature 1.0 equals to the model temperature 0.3.

Prompts for File Uploading and Web Search

For file uploading, please follow the template to create prompts, where {file_name}, {file_content} and {question} are arguments.

file_template = \
"""[file name]: {file_name}
[file content begin]
{file_content}
[file content end]
{question}"""

For Web Search, {search_results}, {cur_date}, and {question} are arguments.

For Chinese query, we use the prompt:

search_answer_zh_template = \
'''# 以下内容是基于用户发送的消息的搜索结果:
{search_results}
在我给你的搜索结果中，每个结果都是[webpage X begin]...[webpage X end]格式的，X代表每篇文章的数字索引。请在适当的情况下在句子末尾引用上下文。请按照引用编号[citation:X]的格式在答案中对应部分引用上下文。如果一句话源自多个上下文，请列出所有相关的引用编号，例如[citation:3][citation:5]，切记不要将引用集中在最后返回引用编号，而是在答案对应部分列出。
在回答时，请注意以下几点：
- 今天是{cur_date}。
- 并非搜索结果的所有内容都与用户的问题密切相关，你需要结合问题，对搜索结果进行甄别、筛选。
- 对于列举类的问题（如列举所有航班信息），尽量将答案控制在10个要点以内，并告诉用户可以查看搜索来源、获得完整信息。优先提供信息完整、最相关的列举项；如非必要，不要主动告诉用户搜索结果未提供的内容。
- 对于创作类的问题（如写论文），请务必在正文的段落中引用对应的参考编号，例如[citation:3][citation:5]，不能只在文章末尾引用。你需要解读并概括用户的题目要求，选择合适的格式，充分利用搜索结果并抽取重要信息，生成符合用户要求、极具思想深度、富有创造力与专业性的答案。你的创作篇幅需要尽可能延长，对于每一个要点的论述要推测用户的意图，给出尽可能多角度的回答要点，且务必信息量大、论述详尽。
- 如果回答很长，请尽量结构化、分段落总结。如果需要分点作答，尽量控制在5个点以内，并合并相关的内容。
- 对于客观类的问答，如果问题的答案非常简短，可以适当补充一到两句相关信息，以丰富内容。
- 你需要根据用户要求和回答内容选择合适、美观的回答格式，确保可读性强。
- 你的回答应该综合多个相关网页来回答，不能重复引用一个网页。
- 除非用户要求，否则你回答的语言需要和用户提问的语言保持一致。

# 用户消息为：
{question}'''

For English query, we use the prompt:

search_answer_en_template = \
'''# The following contents are the search results related to the user's message:
{search_results}
In the search results I provide to you, each result is formatted as [webpage X begin]...[webpage X end], where X represents the numerical index of each article. Please cite the context at the end of the relevant sentence when appropriate. Use the citation format [citation:X] in the corresponding part of your answer. If a sentence is derived from multiple contexts, list all relevant citation numbers, such as [citation:3][citation:5]. Be sure not to cluster all citations at the end; instead, include them in the corresponding parts of the answer.
When responding, please keep the following points in mind:
- Today is {cur_date}.
- Not all content in the search results is closely related to the user's question. You need to evaluate and filter the search results based on the question.
- For listing-type questions (e.g., listing all flight information), try to limit the answer to 10 key points and inform the user that they can refer to the search sources for complete information. Prioritize providing the most complete and relevant items in the list. Avoid mentioning content not provided in the search results unless necessary.
- For creative tasks (e.g., writing an essay), ensure that references are cited within the body of the text, such as [citation:3][citation:5], rather than only at the end of the text. You need to interpret and summarize the user's requirements, choose an appropriate format, fully utilize the search results, extract key information, and generate an answer that is insightful, creative, and professional. Extend the length of your response as much as possible, addressing each point in detail and from multiple perspectives, ensuring the content is rich and thorough.
- If the response is lengthy, structure it well and summarize it in paragraphs. If a point-by-point format is needed, try to limit it to 5 points and merge related content.
- For objective Q&A, if the answer is very brief, you may add one or two related sentences to enrich the content.
- Choose an appropriate and visually appealing format for your response based on the user's requirements and the content of the answer, ensuring strong readability.
- Your answer should synthesize information from multiple relevant webpages and avoid repeatedly citing the same webpage.
- Unless the user requests otherwise, your response should be in the same language as the user's question.

# The user's message is:
{question}'''

How to Run Locally

The model structure of DeepSeek-V3-0324 is exactly the same as DeepSeek-V3. Please visit DeepSeek-V3 repo for more information about running this model locally.

This model supports features such as function calling, JSON output, and FIM completion. For instructions on how to construct prompts to use these features, please refer to DeepSeek-V2.5 repo.

NOTE: Hugging Face's Transformers has not been directly supported yet.

License

This repository and the model weights are licensed under the MIT License.

Citation

@misc{deepseekai2024deepseekv3technicalreport,
      title={DeepSeek-V3 Technical Report}, 
      author={DeepSeek-AI},
      year={2024},
      eprint={2412.19437},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2412.19437}, 
}

Contact

If you have any questions, please raise an issue or contact us at [email protected].