We have released candidate supervised finetuning (SFT) models using both Pythia and LLaMa, as well as candidate reward models for reinforcement learning from human feedback training using Pythia, which you can try, and are beginning the process of applying (RLHF). We have also released the first version of the OpenAssistant Conversations dataset here.

You can play with our best candidate model here and provide thumbs up/down responses to help us improve the model in future!

The candidate Pythia SFT models are available on HuggingFace and can be loaded via the HuggingFace Transformers library. As such you may be able to use them with sufficient hardware. There are also spaces on HF which can be used to chat with the OA candidate without your own hardware. However, these models are not final and can produce poor or undesirable outputs.

LLaMa SFT models cannot be released directly due to Meta's license but XOR weights will be released soon.

The docker compose command in the README is for setting up the project for local development on the website or data collection backend. It does not launch an AI model or the inference server. There is likely no point in running the inference setup and UI locally unless you wish to assist in development.

All Open Assistant code is licensed under Apache 2.0. This means it is available for a wide range of uses including commercial use.

The Open Assistant Pythia based models are released as full weights and will be licensed under the Apache 2.0 license.

The Open Assistant LLaMa based models will be released only as delta weights meaning you will need the original LLaMa weights to use them, and the license restrictions will therefore be those placed on the LLaMa weights.

The Open Assistant data is released under a Creative Commons license allowing a wide range of uses including commercial use.

Open Assistant is a project organized by LAION and developed by a team of volunteers worldwide. You can see an incomplete list of developers on our website.

The project would not be possible without the many volunteers who have spent time contributing both to data collection and to the development process. Thank you to everyone who has taken part!

The model code, weights, and data are free. We are additionally hosting a free public instance of our best current model for as long as we can thanks to compute donation from Stability AI via LAION!

The current smallest (Pythia) model is 12B parameters and is challenging to run on consumer hardware, but can run on a single professional GPU. In future there may be smaller models and we hope to make progress on methods like integer quantisation which can help run the model on smaller hardware.

If you want to help in the data collection for training the model, go to the website https://open-assistant.io/.

If you want to contribute code, take a look at the tasks in GitHub and comment on an issue stating your wish to be assigned. You can also take a look at this contributing guide.

The Python backend for the data collection app as well as for the inference backend uses FastAPI. The frontend is built with NextJS and Typescript.

The ML codebase is largely PyTorch-based and uses HuggingFace Transformers as well as accelerate, DeepSpeed, bitsandbytes, NLTK, and other libraries.

No, it is against their terms of service to use it to help train other models. See this issue. ChatGPT-like answers will be removed.

Yes, you should try. If you are not sure, skip the task.

In your account settings.

You can see a regularly updated interface at https://open-assistant.io/stats.

Check the guidelines.

In the GitHub issues.

We want to ensure that the Open Assistant dataset is as accessible as possible. As such, it's necessary to avoid any harmful or offensive content that could be grounds for removal on sites such as Hugging Face. Likewise, we want the model to be trained to reject as few questions as possible, so it's important to not include prompts that leave the assistant with no other choice but to refuse in order to avoid the generation of harmful content.

If you are using docker-compose instead of docker compose (note the " " instead of the "-"), you should update your docker cli to the latest version. docker compose is the most recent version and should be used instead of docker-compose.

For more details and information check out this StackOverflow thread that explains it all in detail.

BuildKit is Docker's new and improved builder backend. In addition to being faster and more efficient, it supports many new features, among which is the ability to provide a persistent cache, which outlives builds, to compilers and package managers. This is very useful to speed up consecutive builds, and is used by some container images of OpenAssistant's stack.

The BuildKit backend is used by default by Compose V2 (see above).
But if you want to build an image with docker build instead of docker compose build, you might need to enable BuildKit.

To do so, just add DOCKER_BUILDKIT=1 to your environment.

For instance:

export DOCKER_BUILDKIT=1

You could also, more conveniently, enable BuildKit by default, or use Docker Buildx.

We are using pre-commit to ensure the quality of the code as well as the same code standard.

The steps that you need to follow to be able to use it are:

# install the pre-commit Python package
pip3 install pre-commit

# install pre-commit to the Git repo to run automatically on commit
pre-commit install

So from now on, in your next commits it will run the pre-commit on the files that have been staged. Most formatting issues are automatically resolved by the hooks so the files can simply be re-added and you can commit. Some issues may require manual resolution.

If you wish to run pre-commit on all files, not just ones your last commit has modified, you can use pre-commit run --all-files.

Instead of running docker with the root command always, you could create a docker group with granted permissions (root):

# Create new linux user
sudo groupadd docker

# Add the actual user to the group
sudo usermod -aG docker $USER

# Log in the group (apply the group changes to actual terminal session)
newgrp docker

After that, you should be able to run docker: docker run .. In the case you still are not able, can try to reboot terminal:

reboot

If you try to shut down the services (docker-compose down), and you are getting permission denied (using root user), you can try the following:

# Restart docker daemon
sudo systemctl restart docker.socket docker.service

# And remove the container
docker rm -f <container id>

Oftentimes people already have some Postgres instance running on the dev machine. To avoid port problems, change the ports in the docker-compose.yml to ones excluding 5433, like:

1. Change db.ports to - 5431:5431.
2. Add POSTGRES_PORT: 5431 to db.environment
3. Change webdb.ports to - 5432:5431
4. Add POSTGRES_PORT: 5431 to db.environment
5. Add - POSTGRES_PORT=5432 to backend.environment
6. Change web.environment.DATABASE_URL to postgres://postgres:postgres@webdb:5432/oasst_web