📝 NOTE: OpenAssistant is completed, and the project is now finished. Thank you to everyone who contributed! Check out our blog post for more information. The final published oasst2 dataset can be found on HuggingFace at OpenAssistant/oasst2
Open Assistant is a project meant to give everyone access to a great chat based large language model.
We believe that by doing this we will create a revolution in innovation in language. In the same way that stable-diffusion helped the world make art and images in new ways we hope Open Assistant can help improve the world by improving language itself.
The chat frontend is now live here. Log in and start chatting! Please try to react with a thumbs up or down for the assistant's responses when chatting.
The data collection frontend is now live here. Log in and start taking on tasks! We want to collect a high volume of quality data. By submitting, ranking, and labelling model prompts and responses you will be directly helping to improve the capabilities of Open Assistant.
You do not need to run the project locally unless you are contributing to the development process. The website link above will take you to the public website where you can use the data collection app and the chat.
If you would like to run the data collection app locally for development, you can set up an entire stack needed to run Open-Assistant, including the website, backend, and associated dependent services, with Docker.
To start the demo, run this in the root directory of the repository (check this FAQ if you have problems):
docker compose --profile ci up --build --attach-dependencies
Note: when running on MacOS with an M1 chip you have to use:
DB_PLATFORM=linux/x86_64 docker compose ...
Then, navigate to https://localhost:3000
(It may take some time to boot up) and
interact with the website.
Note: If an issue occurs with the build, please head to the FAQ and check out the entries about Docker.
Note: When logging in via email, navigate to
https://localhost:1080
to get the magic email login link.
Note: If you would like to run this in a standardized development environment (a "devcontainer") using vscode locally or in a web browser using GitHub Codespaces, you can use the provided
.devcontainer
folder.
You do not need to run the project locally unless you are contributing to the development process. The website link above will take you to the public website where you can use the data collection app and the chat.
Also note that the local setup is only for development and is not meant to be used as a local chatbot, unless you know what you are doing.
If you do know what you are doing, then see the inference
folder for getting
the inference system up and running, or have a look at --profile inference
in
addition to --profile ci
in the above command.
We are not going to stop at replicating ChatGPT. We want to build the assistant of the future, able to not only write email and cover letters, but do meaningful work, use APIs, dynamically research information, and much more, with the ability to be personalized and extended by anyone. And we want to do this in a way that is open and accessible, which means we must not only build a great assistant, but also make it small and efficient enough to run on consumer hardware.
We want to get to an initial MVP as fast as possible, by following the 3-steps outlined in the InstructGPT paper
- Collect high-quality human generated Instruction-Fulfillment samples (prompt + response), goal >50k. We design a crowdsourced process to collect and reviewed prompts. We do not want to train on flooding/toxic/spam/junk/personal information data. We will have a leaderboard to motivate the community that shows progress and the most active users. Swag will be given to the top-contributors.
- For each of the collected prompts we will sample multiple completions. Completions of one prompt will then be shown randomly to users to rank them from best to worst. Again this should happen crowd-sourced, e.g. we need to deal with unreliable potentially malicious users. At least multiple votes by independent users have to be collected to measure the overall agreement. The gathered ranking-data will be used to train a reward model.
- Now follows the RLHF training phase based on the prompts and the reward model.
We can then take the resulting model and continue with completion sampling step 2 for a next iteration.
All open source projects begin with people like you. Open source is the belief that if we collaborate we can together gift our knowledge and technology to the world for the benefit of humanity.
Check out our contributing guide to get started.