Still worried about not being able to play ChatGPT? You can also experience intelligent dialogue with these dozen open source alternatives

It has been almost four months since the initial release of ChatGPT. During this period, the amazing performance of ChatGPT made people wonder whether AIGC will be realized in advance.

However, a well-known secret is that ChatGPT is unlikely to be open source. Coupled with the huge investment in computing power and massive training data, it has set many obstacles for the research community to copy its implementation process. level.

Faced with the onslaught of ChatGPT, open source alternatives are a good choice. While they achieve similar functions, they can also allow you to understand the operating mechanism of ChatGPT through alternatives.

With the advancement of technology, many other new AI models similar to ChatGPT have now emerged. In this article, we will take stock of ChatGPT’s open source replacement projects. The content will be divided into two parts, one is a summary of ChatGPT-like open source projects, and the other is a supplement to the project summary.

Project summary

The project author nichtdax has compiled 9 open source ChatGPT replacement projects. They will be introduced one by one next.

Project address: https://github.com/nichtdax/awesome -totally-open-chatgpt

PaLM-rlhf-pytorch​

The first project is "PaLM-rlhf-pytorch", the project author is Phil Wang. This project is an implementation of RLHF (Reinforcement Learning with Human Feedback) on top of the PaLM architecture, which is basically ChatGPT using PaLM.

This project has received 5.8k Stars on GitHub.

Project address: https://github.com/lucidrains/PaLM-rlhf- pytorch

#The picture below shows the training process.

OpenChatKit

The second project is "OpenChatKit", which Provides a powerful open source foundation for creating specialized and general-purpose chatbots for a variety of applications. The kit includes an instruction-tuned 20 billion parameter language model, a 6 billion parameter tuning model, and a scalable retrieval system that covers the latest responses from a custom repository.

OpenChatKit is trained on the OIG-43M training dataset, which is a union of Together, LAION, and Ontocord.ai. The project author stated that this is not only the release of a model, but also the beginning of an open source project. They are releasing a set of tools and processes and continually improving them through community contributions.

The project has received 5.7k Stars on GitHub.

## Project address: https://github.com/togethercomputer/OpenChatKit

text-generation-webui

The third project is "text-generation-webui", which is a tool for running GPT-J 6B, OPT, GALACTICA, LLaMA and Gradio web UI for large language models such as Pygmalion. This project aims to be the AUTOMATIC1111/stable-diffusion-webui of text generation.

Features include using drop-down menus to switch between different models, providing a notebook mode similar to OpenAI playground, chat mode for dialogue and role-playing, generating beautiful HTML output and more.

The project has received 3.4k on GitHub.

Project address: https://github.com/oobabooga/text-generation- webui

KoboldAI-Client

The fourth project is "KoboldAI-Client", which is a browser-based front end through multiple local and remote AI models Implement AI-assisted writing.

KoboldAI-Client provides a standard set of tools, including memory, author notes, world information, save & load, adjustable AI settings, formatting options, and importing existing text The abilities of the adventure game AI Dungeon. You can turn on Adventure mode or play games like AI Dungeon Unleashed.

The project has received 1.4k on GitHub.

## Project address: https://github.com/KoboldAI/KoboldAI-Client

Open-Assistant

The fifth project is "Open-Assistant", which aims to make large chat-based language models accessible to everyone. The project authors hope to spark a revolution in language innovation, just as stable diffusion helps the world create art and images in new ways.

The project author plans to collect high-quality manually generated instruction execution samples (instruction responses), with a goal of greater than 50k. For each indication collected, they will sample multiple completion results. Next, enter the RLHF training phase based on the instruction and reward model.

The project has received 19k stars on GitHub.

## Project address: https://github.com/LAION-AI/Open- Assistant

stanford_alpaca

The sixth project is "stanford_alpaca", which aims to build and share an instruction-following LLaMA model. This repo contains 52k data for fine-tuning the model, code for generating the data, and code for fine-tuning the model.

The project has received 9.5k Stars on GitHub.

## Project address: https://github.com/tatsu-lab/stanford_alpaca

ChatRWKV

The seventh project is "ChatRWKV", which is similar to ChatGPT, but is supported by the RWKV (100% RNN) model and is open source. According to the project authors, RWKV is currently the only RNN model that can match transformers in terms of quality and scalability, while being faster and saving VRAM.

The project has received 3.5k Stars on GitHub.

## Project address: https://github.com/BlinkDL/ChatRWKV

ChatGLM-6B

The eighth project is the company’s Zhipu AI open source, Chinese-English bilingual conversation language model "ChatGLM-6B" transformed from Tsinghua’s technological achievements, based on the General Language Model (GLM) architecture with 6.2 billion parameters. Combined with model quantization technology, users can deploy it locally on consumer-grade graphics cards (a minimum of 6GB of video memory is required at the INT4 quantization level).

ChatGLM-6B uses technology similar to ChatGPT and is optimized for Chinese Q&A and dialogue. After bilingual training in Chinese and English with about 1T identifiers, supplemented by supervised fine-tuning, feedback self-service, human feedback reinforcement learning and other technologies, the 6.2 billion parameter ChatGLM-6B has been able to generate answers that are quite consistent with human preferences.

However, due to the small scale of ChatGLM-6B, it is currently known to have quite a few limitations, such as factual/mathematical logic errors, the possibility of generating harmful/biased content, and Weak contextual skills, confused self-perception, and production of English instructions that are completely contradictory to Chinese instructions.

This project has received 6k stars on GitHub.

## Project address: https://github.com/THUDM/ChatGLM-6B

xmtf

The ninth project is "xmtf", which contains all the components of BLOOMZ, mT0 and xP3 introduced in the creation paper "Crosslingual Generalization through Multitask Finetuning".

Among them, BLOOMZ is a 176 billion parameter and open-access multi-language model; mT0 (here specifically refers to Google's T5X); xP3 is a supervised data set in 46 languages, with There are prompts in English and machine translation.

Project address: https://github.com/bigscience-workshop/xmtf

In addition to the models mentioned above, the ChatGPT replacement project also includes some research based on the Meta open source large model series LLaMA (Large Language Model Meta AI). The parameters of this series of models range from It ranges from 7 billion to 65 billion. The LLaMA model with 13 billion parameters can outperform GPT-3 (175 billion parameters) "on most benchmarks" and can run on a single V100 GPU; while the largest LLaMA model with 65 billion parameters is comparable to Google Chinchilla-70B and PaLM-540B.

• Paper link: https://research.facebook.com/publications/llama-open-and-efficient-foundation-language -models/
• GitHub link: https://github.com/facebookresearch/llama

Based on this research, developer Several open source projects have also been proposed.

llama.cpp No need for GPU, you can run LLaMA

llama.cpp project realizes running LLaMA on MacBook, and developers have successfully run LLaMA on MacBook with 4GB RAM LLaMA 7B is running on the Raspberry Pi. In summary, developers can run LLaMA models even if they don’t have a GPU.

## Project address: https://github.com/ggerganov/llama.cpp

ChatLLaMA​

Since the LLaMA large model series does not use the RLHF method, the startup Nebuly AI open sourced the training method of the RLHF version of LLaMA (ChatLLaMA). Its training process is similar to ChatGPT, and the project allows building ChatGPT-style services based on pre-trained LLaMA models. Compared with ChatGPT, LLaMA has a smaller architecture, but the training process and single-GPU inference are faster and cheaper; the library also supports all LLaMA model architectures (7B, 13B, 33B, 65B), so users can and inference performance preferences to fine-tune the model.

Project address: https://github.com/nebuly-ai/nebullvm/ tree/main/apps/accelerate/chatllama

Looking forward to more open source projects joining in the future.

The above is the detailed content of Still worried about not being able to play ChatGPT? You can also experience intelligent dialogue with these dozen open source alternatives. For more information, please follow other related articles on the PHP Chinese website!