## Huggingface with MiniCPM

Welcome to the guide on using Huggingface with MiniCPM. Below you will find detailed information on the MiniCPM model as well as how to install and use it with Huggingface.

### What is MiniCPM?

MiniCPM is an End-Size LLM developed by ModelBest Inc. and TsinghuaNLP, with only 2.4B parameters excluding embeddings. It has shown close performance compared to Mistral-7B on open-sourced general benchmarks with better ability on Chinese, Mathematics, and Coding after SFT (Secure Fine-Tuning). The overall performance exceeds other large models such as Llama2-13B, MPT-30B, Falcon-40B, etc. MiniCPM-V, based on MiniCPM-2B, achieves the best overall performance among multimodel models of the same scale, surpassing existing multimodal large models built on Phi-2 and achieving performance comparable to or even better than 9.6B Qwen-VL-Chat on some tasks.

### Licensing and Limitations

MiniCPM is released under an open-source license for research and limited commercial use. However, it comes with certain limitations due to its model size and capacity. It is important to review and understand these limitations before using the model.

### Downloading and Usage

To use MiniCPM with Huggingface, it is necessary to install `transformers>=4.36.0` and `accelerate`. After installation, you can use the provided Python code to initiate and interact with the model. Just make sure to specify the data type of the model clearly in the `from_pretrained` method.

The expected output should match the input provided to the model.

### Citation

If you use MiniCPM in your work, it is important to cite it properly. Use the following format for citation:

“`plaintext
@inproceedings{minicpm2024,
title={MiniCPM:Unveiling the Potential of End-side Large Language Models},
booktitle={OpenBMB Blog},
year={2024}
}
“`

### Evaluation and Further Details

For more detailed evaluation and specific details on using MiniCPM with Huggingface, you can refer to the official GitHub repository containing the readme and evaluation results.

It is important to note that the quality of Huggingface generation has been observed to be slightly lower than vLLM. Therefore, benchmarking using vLLM is recommended.

### Conclusion

By following this guide, you should be able to understand, install, and utilize MiniCPM with Huggingface for your language model needs. Remember to keep the limitations in mind and cite the model properly when using it in your work.

Source link
## Huggingface Tutorial

Welcome to the tutorial on how to use MiniCPM with Huggingface! In this tutorial, we will walk you through the process of downloading the model and using it for language generation.

### Model Download
First, you’ll need to download the MiniCPM model. The model is available for academic research and limited commercial use. You can download the model from the following link: [Download MiniCPM](https://github.com/OpenBMB/MiniCPM/blob/main/README-en.md#model-download).

### Model Usage
Once you have downloaded the model, you can use it for language generation. Make sure you have `transformers>=4.36.0` and `accelerate` installed before running the following code:

“`python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

torch.manual_seed(0)

path = ‘openbmb/MiniCPM-2B-sft-bf16′
tokenizer = AutoTokenizer.from_pretrained(path)
model = AutoModelForCausalLM.from_pretrained(path, torch_dtype=torch.bfloat16, device_map=’cuda’, trust_remote_code=True)

responds, history = model.chat(tokenizer, “Enter your prompt here”, temperature=0.8, top_p=0.8)
print(responds)
“`

Ensure that you specify the model’s data type and device in the `from_pretrained` call to avoid calculation errors.

Expected output should resemble the following:
“`
Sample output here
“`

### License and Usage Statement
Please note that as a language model, MiniCPM generates content based on learned patterns and does not represent the viewpoints of the model developers. Therefore, users are responsible for evaluating and verifying the generated content.

Additionally, if any issues arise from the use of MiniCPM, including but not limited to data security issues or public opinion risks, the developers will not be held responsible.

### Citation
If you utilize MiniCPM in your work, please provide the following citation:
“`
@inproceedings{minicpm2024,
title={MiniCPM:Unveiling the Potential of End-side Large Language Models},
booktitle={OpenBMB Blog},
year={2024}
}
“`

Get started with MiniCPM and have fun exploring the potential of large language models with Huggingface!

MiniCPM 技术报告 Technical Report |
OmniLMM 多模态模型 Multi-modal Model |
CPM-C 千亿模型试用 ~100B Model Trial

MiniCPM 是面壁与清华大学自然语言处理实验室共同开源的系列端侧语言大模型,主体语言模型 MiniCPM-2B 仅有 24亿(2.4B)的非词嵌入参数量。

  • 经过 SFT 后,MiniCPM 在公开综合性评测集上,MiniCPM 与 Mistral-7B相近(中文、数学、代码能力更优),整体性能超越 Llama2-13B、MPT-30B、Falcon-40B 等模型。
  • 经过 DPO 后,MiniCPM 在当前最接近用户体感的评测集 MTBench上,MiniCPM-2B 也超越了 Llama2-70B-Chat、Vicuna-33B、Mistral-7B-Instruct-v0.1、Zephyr-7B-alpha 等众多代表性开源大模型。
  • 以 MiniCPM-2B 为基础构建端侧多模态大模型 MiniCPM-V,整体性能在同规模模型中实现最佳,超越基于 Phi-2 构建的现有多模态大模型,在部分评测集上达到与 9.6B Qwen-VL-Chat 相当甚至更好的性能。
  • 经过 Int4 量化后,MiniCPM 可在手机上进行部署推理,流式输出速度略高于人类说话速度。MiniCPM-V 也首次跑通了多模态大模型在手机上的部署。
  • 一张1080/2080可高效参数微调,一张3090/4090可全参数微调,一台机器可持续训练 MiniCPM,二次开发成本较低。

我们将完全开源MiniCPM-2B的模型参数供学术研究和有限商用,以及训练过程中的所有Checkpoint和大部分非专有数据供模型机理研究。

  • 基于MiniCPM-2B的指令微调与人类偏好对MiniCPM-2B-SFT/DPO。
  • 基于MiniCPM-2B的多模态模型MiniCPM-V,能力超越基于Phi-2的同参数级别多模态模型
  • MiniCPM-2B-SFT/DPO的Int4量化版MiniCPM-2B-SFT/DPO-Int4。
  • 基于MLC-LLM、LLMFarm开发的MiniCPM手机端程序,文本及多模态模型均可在手机端进行推理。

MiniCPM is an End-Size LLM developed by ModelBest Inc. and TsinghuaNLP, with only 2.4B parameters excluding embeddings.

  • MiniCPM has very close performance compared with Mistral-7B on open-sourced general benchmarks with better ability on Chinese, Mathmetics and Coding after SFT. The overall performance exceeds Llama2-13B, MPT-30B, Falcon-40B, etc.
  • After DPO, MiniCPM outperforms Llama2-70B-Chat, Vicuna-33B, Mistral-7B-Instruct-v0.1, Zephyr-7B-alpha, etc. on MTBench.
  • MiniCPM-V, based on MiniCPM-2B, achieves the best overall performance among multimodel models of the same scale, surpassing existing multimodal large models built on Phi-2 and achieving performance comparable to or even better than 9.6B Qwen-VL-Chat on some tasks.
  • MiniCPM can be deployed and infer on smartphones, and the speed of streaming output is relatively higher than the verbal speed of human. MiniCPM-V is the first multi-modal models that can be deployed on smartphones.
  • The cost of developing based on MiniCPM is low. Parameter efficient finetuning can be conducted with a single 1080/2080 GPU and full parameter finetuning can be conducted with a 3090/4090 GPU.

We release all model parameters for research and limited commercial use. We also release all the checkpoint during training and most public training data for research on model mechanism.

  • SFT and DPO version based on MiniCPM-2B and human preference: MiniCPM-2B-SFT/DPO
  • The multi-modal model MiniCPM-V based on MiniCPM-2B, which outperforms models with similar size, i.e., Phi-2
  • The INT4 quantized version MiniCPM-2B-SFT/DPO-Int4 based on MiniCPM-2B-SFT/DPO
  • Mobile phone application based on MLC-LLM and LLMFarm. Both language model and multimodel model can conduct inference on smartphones.



评测结果 Evaluation Results

详细的评测结果位于github仓库

Detailed evaluation results are in github repo

注意:我们发现使用Huggingface生成质量略差于vLLM,因此推荐使用vLLM进行测试。我们正在排查原因。

Notice: We discovered that the quality of Huggingface generation is slightly lower than vLLM, thus benchmarking using vLLM is recommended.
We are investigating the cause now.



局限性 Limitations

  • 受限于模型规模,模型可能出现幻觉性问题。其中由于DPO模型生成的回复内容更长,更容易出现幻觉。我们也将持续进行MiniCPM模型的迭代改进;

  • 为了保证在学术研究用途上模型的通用性,我们未对模型进行任何身份认同训练。同时由于我们用ShareGPT开源语料作为部分训练数据,模型可能会输出类似GPT系列模型的身份认同信息;

  • 受限于模型规模,模型的输出受到提示词(prompt)的影响较大,可能多次尝试产生不一致的结果;

  • 受限于模型容量,模型的知识记忆较不准确,后续我们将结合RAG方法来增强模型的知识记忆能力。

  • Due to limitations in model size, the model may experience hallucinatory issues. As DPO model tend to generate longer response, hallucinations are more likely to occur. We will also continue to iterate and improve the MiniCPM model.

  • To ensure the universality of the model for academic research purposes, we did not conduct any identity training on the model. Meanwhile, as we use ShareGPT open-source corpus as part of the training data, the model may output identity information similar to the GPT series models.

  • Due to the limitation of model size, the output of the model is greatly influenced by prompt words, which may result in inconsistent results from multiple attempts.

  • Due to limited model capacity, the model’s knowledge memory is not accurate. In the future, we will combine the RAG method to enhance the model’s knowledge memory ability.



模型下载 Download



模型使用 Usage

  • 安装transformers>=4.36.0以及accelerate后,运行以下代码
  • 注意:需要在from_pretrained中明确指明模型的数据类型,否则会引起较大计算误差
  • Run the following code after install transformers>=4.36.0 and accelerate
  • Warning: It is necessary to specify the data type of the model clearly in ‘from_pretrained’, otherwise large calculation errors will be caused
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
torch.manual_seed(0)

path = 'openbmb/MiniCPM-2B-sft-bf16'
tokenizer = AutoTokenizer.from_pretrained(path)
model = AutoModelForCausalLM.from_pretrained(path, torch_dtype=torch.bfloat16, device_map='cuda', trust_remote_code=True)

responds, history = model.chat(tokenizer, "山东省最高的山是哪座山, 它比黄山高还是矮?差距多少?", temperature=0.8, top_p=0.8)
print(responds)
  • 期望输出 Expected Output
山东省最高的山是泰山,海拔1545米。

相对于黄山(海拔1864米),泰山海拔较低,相差约319米。



开源协议 LICENSE



模型协议 Model LICENSE



声明 Statement

  • 作为一个语言模型,MiniCPM 通过学习大量的文本来生成内容,但它无法理解、表达个人观点或价值判断,它所输出的任何内容都不代表模型开发者的观点和立场。

  • 因此用户在使用 MiniCPM 生成的内容时,应自行负责对其进行评估和验证。

  • 如果由于使用 MinCPM 开源模型而导致的任何问题,包括但不限于数据安全问题、公共舆论风险,或模型被误导、滥用、传播或不当利用所带来的任何风险和问题,我们将不承担任何责任。

  • As a language model, MiniCPM generates content by learning from a vast amount of text.

  • However, it does not possess the ability to comprehend or express personal opinions or value judgments.

  • Any content generated by MiniCPM does not represent the viewpoints or positions of the model developers.

  • Therefore, when using content generated by MiniCPM, users should take full responsibility for evaluating and verifying it on their own.



工作引用 Citation

@inproceedings{minicpm2024,
 title={MiniCPM:Unveiling the Potential of End-side Large Language Models},
 booktitle={OpenBMB Blog},
 year={2024}
}

The `

` tag in HTML is one of the most versatile and widely used elements for creating structure and organizing content on a webpage. It is a block-level element that can be used to group together and style other HTML elements. Here are some use cases for the `

` tag:

1. Creating Sections: The `

` tag is often used to create sections of a webpage, such as a header, footer, sidebar, or main content area. By using `

` elements with different classes or IDs, web developers can style and position these sections as needed.

2. Layout and Positioning: `

` elements are often used in conjunction with CSS to create complex layouts and positioning of elements on a webpage. By using CSS properties such as `display`, `position`, and `float`, developers can achieve the desired layout and positioning of content.

3. Grouping Content: The `

` tag can be used to group together related content, such as a group of images, a set of links, or a block of text. This allows developers to apply styles and effects to the entire group of content as a single unit.

4. Creating Containers: `

` elements are commonly used to create containers for content, such as a card or a modal. By enclosing content within a `

`, developers can apply background colors, borders, padding, and other styling to create visually appealing containers.

5. Responsive Design: When building responsive webpages, developers often use `

` elements to create flexible, grid-based layouts. By using CSS frameworks or custom styles, developers can create responsive designs that adapt to different screen sizes and devices.

6. JavaScript and DOM Manipulation: The `

` tag is often targeted by JavaScript and used for dynamic content manipulation. For example, developers can use JavaScript to dynamically add, remove, or modify the content within a `

` element based on user interactions.

7. Accessibility: `

` elements can be used to improve the accessibility of a webpage by providing semantic structure to the content. Developers can use `

` elements in combination with ARIA roles and attributes to enhance the accessibility of complex web components.

In summary, the `

` tag in HTML is a fundamental building block for creating structure, layout, and organization on a webpage. It is a versatile and powerful element that is used in a wide range of use cases, from basic content grouping to complex responsive layouts and dynamic content manipulation.

2024-02-02T12:38:26+01:00