Hello everyone, my name is Jack.
Today is Qingming, where did you guys go to play?
I have already gone out for a while last month, and I honestly stayed at home to rest in Qingming.
After looking through some recent hot open source projects, I found that AIGC is still the world.
Today, I will continue to focus on selecting a few recent high-quality open source projects. I believe these projects will be helpful to our daily work, study and life.
Today I organize and share with you, hoping to inspire you.
ChatDoctor
I believe that there will be more and more GPT large models for specific domains.
A very professional GPT large model has been trained in various fields. At that time, with the addition of an upper-level scheduling module, an all-round AI in various fields will appear.
ChatDoctor is an attempt in the medical field, which is based on the fine-tune of Meta's open source LLaMA.
ChatDoctor: What's wrong with you?
The model has undergone multiple rounds of fine-tuning:
-
First use the 52K instruction-following data provided by Stanford Alpaca to train a general dialogue model;
-
Fine-tune the model again through the "5K conversation dataset between patients and doctors" generated by ChatGPT GenMedGPT-5k and disease database;
-
Third round of fine-tuning through real conversations between patients and physicians at HealthCareMagic-200k;
-
Fourth round of fine-tuning through real conversations between patients and physicians of icliniq-26k.
It is the developer's vision to use these tailored doctor-patient conversations to fine-tune LLMs, and the resulting models have great potential to understand patients' needs, provide informed advice, and provide valuable s help.
project address:
https://github.com/Kent0n-Li/ChatDoctor
In addition to this open source ChatDoctor, there is also BloombergGPT, a large model in the financial world.
BloombergGPT was created by Bloomberg, the world's largest financial information company. During the 40 years since its establishment, the company has collected a large amount of financial market data. For financial scenarios, Bloomberg, which has huge advantages, is naturally unwilling to lag behind, so it developed this 50 billion parameters. The large language model BloombergGPT.
But just take a look at this. If the effect of BloombergGPT is really good, 99.99% will not be open source, nor will financial data. After all, this is the real "technical barrier".
Yes
The Chinese name is Bai Ze, and he claims that he can train his own chatbot in a few hours through chatgpt, and only needs a GPU graphics card.
The current open source models are:
-
Baize-7b
-
Baize-13B
-
Baize-30B
-
Baize Healthcare-7B
-
Baize Chinese-7B (Coming soon)
The Chinese version of the model will soon be open-sourced, and this project is worth looking forward to.
project address:
https://github.com/project-baize/baize-chatbot
tomesd
tomesd is still very useful. A brief summary of its function is: acceleration.
Using tomesd can greatly increase the speed of stable diffusion to generate images on the basis of ensuring the quality of image generation.
FID is an evaluation index of image generation quality. It can be seen that after using tomesd, SD image generation speed has increased by more than 2 times, memory usage has decreased by more than 5 times, and image generation quality is still very high. This is also considered a large-scale generation of images. It is a sharp tool, saving a lot of GPU resources and memory resources.
project address:
https://github.com/dbolya/tomesd
JARVIS
Yes, its Chinese name is Jarvis, and it is the upper-level scheduling system I mentioned above.
It can be based on your input, such as "Please generate a picture of a girl reading a book, her posture is similar to the example picture I provided. Then use your voice to describe this picture."
This scheduling system will automatically analyze which AI models to call based on your input, realize your requirements, and directly call the corresponding models on HuggingFace to help you complete the execution.
Well, now the work of "Switching Man" is also done by AI for you.
project address:
https://github.com/microsoft/JARVIS
Text2Video-Zero
Text2Video, as the name suggests, uses text descriptions to generate videos.
For example, enter: a panda, playing the guitar in Times Square.
Another example is input: a cute kitten running on the lawn.
The project has also been deployed on Hugging Face, and those who want to experience it can directly and retrograde.
project address:
https://github.com/picsart-ai-research/text2video-zero
https://huggingface.co/spaces/PAIR/Text2Video-Zero
at last
At the end of the article, let’s talk about what just happened: ChatGPT stopped selling Plus members.
The Plus upgrade button on the official website is grayed out.
Some netizens said that this is a kind of marketing method, which keeps the heat of reports from all parties.
I also applied for the API interface of GPT 4.0 before, but it has not passed yet.
In addition, the incident of ChatGPT being jointly banned has intensified. It seems that all of a sudden, the whole world has ignited a momentum to siege ChatGPT.
In the previous joint open letter of the big guys, the number of signers has risen to more than 9,000.
https://futureoflife.org/open-letter/pause-giant-ai-experiments/
In the past two days, I launched a wave of research on GPT-4 whether it is necessary to suspend the voting for 6 months. This is the result of our readers' voting:
Let's wait and see where the development of AI large models will go.