OpenAssistant: The Ultimate Open-Source Chat AI Experience!

What is Open Assistant?
Open Assistant is a chat-based assistant that understands tasks and can interact with third-party systems. It retrieves information dynamically through the inclusion of plugins and data collection.

Introduction to Open Assistant
Open Assistant is an open-source AI assistant that has been developed by a global community effort and is organized by LAION. Over 600,000 interactions with humans have been collected to build a unique dataset. The dataset contains over 10,000 fully annotated conversation trees on diverse topics ranging from programming to making an omelette. Contributions from over 13,000 volunteers from around the world made it possible to collect data in many languages.

Models and Data Set
First models have already been trained on this data. Team claims that  the models are so powerful and capable that they may be kept private until fully understood for safety reasons. However, they made  everything is available including the data set, code, chat interface and paper. A weights and biases report will be released along with the paper. All items mentioned in this article will be linked in the ‘Source’ section at the end.

Chat Interface & Accessibility
The chat interface is by far the best place to try out these things as it serves as an evaluation for new models. There are llama-based models which are licensed research only but also pythia-based models which are fully open source and business-friendly. Different models can be found under the "Model" tab, with k50 providing straightforward answers and nucleus providing more creative ones. Users can log in with Discord or email, However It seems there are issues with emails ending up in spam.

Open Assistant's Achievements
Members claim that Open Assistant's achievements as a project is its use of human demonstration data. They emphasize that Open Assistant has the largest collection of human demonstration data of its kind. They explain that humans are more creative and resourceful than synthetic or self-instructed data. They even clam that Open Assistant's data may even be better than open ai's because they have to pay crowd workers for their data.

Future Plans
Members mention that there are users who are investigating plugins for Open Assistant. These plugins allow users to interact with other services like Google search.

Comparison with Chat GPT
A survey was conducted where users rated both an Open Assistant model based on Pythia and Chat GPT. Results showed that preference was dead even at 48.3% for Open Assistant and 51.7% for Chat GPT.

Advantages of Open Assistant
While not as good as open AI models in coding, Open Assistant models are more human, concise, interesting, and fun to interact with.

Successes of Data Collection
This includes high-quality contributions from users who reported enjoyment contributing to their first open-source project. Spam removal was found to be extremely effective at removing ill-suited contributions while making optimal use of valuable time from human moderators who were volunteers. Many people have contributed to the data code, documentation, data moderation, and training. Almost 250 people have contributed to the GitHub repository.

Hugging Face Streaming Support
The text generation inference server comes from this feature and a similar thing can be found in the main Hugging Face library.



sources
Chat UI: https://open-assistant.io/chat
Website Link: https://open-assistant.io Dataset: https://huggingface.co/datasets/OpenAssistant/oasst1 Code: https://github.com/LAION-AI/Open-Assistant Paper (temporary): https://ykilcher.com/oa-paper