Chatbots and Language Model Playgrounds

Ai.ra has a new look, with several improvements 🤖

Ai.ra is AIRES at PUCRS’s expert chatbot. Developed by Nicholas Kluge and Carolina Del Pino, Ai.ra is open software that can be accessed in the following repository.

Ai.ra has four iterations. The first three were trained through machine learning (a Bayesian network, a bidirectional LSTM, and a Transformer, all trained through supervised learning), while the fourth was built from predefined rules (n-gram analysis + dictionary search). Currently, Ai.ra is able to answer a range of concept-related questions (disambiguation) in areas such as Machine Learning, Ethics, and AI Safety.

In our online version of Ai.ra, you can also chat with a fine-tuned/distilled version of BlenderBot 3, a large language model developed by Meta AI. People interested in testing and evaluating such models can use our UI at the following link. This tool was built for research and educational purposes only.

AIRES Alignment Playground 🎯

Another tool we’re now making available is the Alignment Playground, a purpose-built UI that lets people interact with and test the alignment capabilities of large language models. Our playground is currently open, and was designed to let people without a technical background interact and experiment with such systems. The model we currently use is Bloom, one of the largest open-source language models ever built.

“Alignment is about the ability of ML models to robustly follow human instructions, and to understand the intentions of a human controller.”

Our playground’s implementation can be found in our repository, Teeny-Tiny Castle 🏰. We hope these tools can help spread knowledge and help researchers investigate the limitations and issues related to the use of language models in conversational and text-generation tasks.

🤗