UC Berkeley Researchers Introduce Koala: A New AI Chatbot That Tweaks Dialogue Near ChatGPT Quality

Systems such as ChatGPT, Bard, Bing Chat, and Claude respond to various user queries, provide sample code, and can even create poems thanks to their large scale language model (LLM).

The most powerful LLMs typically require large compute resources for training, and therefore must use large private datasets. Open-source models are probably not as powerful as closed-source models, but with the right training data, you might get close. Smaller open-source models can be significantly improved with the right data, as evidenced by projects like Alpaca at Stanford University that use OpenAI’s GPT model data to fine-tune LLaMA.

A recent UC Berkeley AI study presents a new model called Koala. Koala is trained using data that includes interactions with competent closed-source models like ChatGPT. This data is available on the web and used in training. It uses online scraped dialogue data, a question-answer dataset, and a human feedback dataset. The researcher fine-tunes his LLaMA base model. The dataset contains high-quality responses to user queries from large existing language models.

Curation of training data is a major obstacle in developing conversational AI. Many existing chat models use custom datasets that require extensive human annotation. Koala’s training set is curated from curated conversational data from the internet and public sources. Conversations between users and large language models (such as ChatGPT) are included in this data set.

Instead of trying to get as much data from the web as possible, the team chose quality over quantity. Question-answering, human feedback (positive and negative ratings), and conversations with existing language models were all conducted using publicly available datasets.

The team conducted a trial to compare two models: one that relied only on distillation data (Koala-Distill) and one that used all available data including distillation data and open source data (Koala-All). The we. They examine how well these models perform and assess the impact of distillation and public datasets on the final results. In their human evaluation, they compared Koala-All to his Koala-Distill, Alpaca, and ChatGPT to test its pace.

The training data for the Alpaca model is contained in the Alpaca test set. This test set consists of representative user prompts taken from the self-study dataset. They also provide a (Koala) test set consisting of 180 real user queries submitted online to provide a more realistic second evaluation process for him. These questions come from a wide range of users and are written in a natural conversational tone. It gives a better indication of how people are using chat-based services. Using these two sets of evaluation data, the researchers asked approximately 100 evaluators to compare the quality of their model outputs for these hidden sets of tasks using the Amazon Mechanical Turk platform. .

Koala-All performed similarly to Alpaca on the Alpaca test set. On the other hand, based on a proposed test set consisting of real customer questions, Koala oar scored better than alpaca in almost half of the cases, and outperformed alpaca in 70% of cases. Iruka was tied with Alpaca.

The team said that dialogue tweaks could cause koalas to hallucinate and make counterfactual comments in a very confident tone. If so, future research should explore potential shortcomings of small models that inherit the confident style of larger language models before inheriting the same level of fact.

This article bear blog koala and its demo. All credit for this research goes to the researchers of this project.Also, don’t forget to participate Our 17k+ ML SubReddit, cacophony channeland email newsletterWe share the latest AI research news, cool AI projects, and more.