We are interested in Unsupervised Domain Adaptation for LLMs.

  • Problem: source and target distribution do not match
  • Sometimes called dataset shift. In NLP, domain shift
  • Not transfer learning, but can be seen as a particular case of it.
  • Transductive transfer learning.
  • We look into unsupervised and task-independant approaches
  • Division into model-centric and data-centric approaches

We presented recent work on SNCF conversational data uses a mix of most of these ideas !

Les slides du séminaire