Microsoft open-sources GODEL language model for real-world conversations

Microsoft open-sources GODEL language model for real-world conversations

Microsoft open-sources GODEL language model for real-world conversations.

Microsoft has announced open source its grounded Open Dialogue Language Model (GODEL), its large-scale pretrained dialogue language model.

According to the introduction, GODEL introduces a new class of pre-trained language models that enable task-oriented dialogue and social dialogue, and are evaluated by the usefulness of their responses.

GODEL’s predecessor, DialogGPT, was Microsoft’s first large-scale pre-trained language model specifically designed for dialogue, released in 2019, to help make conversational AI more accessible and usable.

Compared to DialoGPT, GODEL utilizes a new stage of base pre-training designed to better support fine-tuning stages that require external information from the current dialogue (such as a database or file) to generate a good response.

One of the main features of GODEL is that it provides users with the flexibility to define the basis of their models, which is mainly reflected in GODEL’s versatility in different dialog contexts.

For example, if someone asks about a local restaurant, GODEL will be able to provide a specific and accurate answer, even though that place may not be included in the data used to train it.

Another key feature is its extensive conversational app.

GODEL can be applied to a variety of conversations, including task-oriented, Q&A, and basic chat conversations. In the same conversation, GODEL can produce reasonable responses to various query types, including general questions or requests for specific operations.

Microsoft said that the results of its paper showed that GODEL’s answer was also assessed as helpful. “We compare our model with several robust baselines and state-of-the-art methods, and show that GODEL excels in both human and automated evaluation.”

At present, Microsoft has provided GODEL completely open source; including three different versions: base, large and extra-large.

And provides the code needed to retrain all pretrained models and fine-tune models for specific tasks: CoQA datasets for conversational question answering; Wizard of Wikipedia and Wizard of the Internet datasets for information-seeking chats ; MultiWOZ for Complete the dialogue on the mission.

“We hope GODEL will help numerous academic research teams advance the field of conversational AI with innovative conversational models while eliminating the need for massive GPU resources. We plan to continuously improve GODEL and provide more models to the research community.”