GODEL: Large-Scale Pre-training for Goal-Directed Dialog

This is the home page of project GODEL (Grounded Open Dialogue Language Model), a large open-source pre-trained language model for dialog. In contrast with its predecessor DialoGPT (opens in new tab), GODEL leverages a new phase of grounded pretraining designed to better support finetuning phases that require information external to the current conversation (e.g., a database or document) to produce good responses. Experiments against a benchmark suite combining task-oriented dialog, conversational QA, and grounded open-domain dialog show that GODEL outperforms state-of-the-art pre-trained dialog models in few-shot finetuning setups, in terms of both human and automatic evaluation. A novel feature of the evaluation methodology in GODEL is the introduction of a notion of utility that assesses the usefulness of responses (extrinsic evaluation) in addition to their communicative features (intrinsic evaluation). We show that extrinsic evaluation offers improved inter-annotator agreement and correlation with automated metrics. More information about this work can be found in the paper “GODEL: Large-Scale Pre-training for Goal-Directed Dialog (opens in new tab).”

The code and models of GODEL are available on GitHub (opens in new tab), with three model sizes currently available: base, large, and extra-large. We will post information about new releases and papers on GODEL on this project page.