Fine-tuned response generation

[RyanSelesnik]

Could potentially fine-tune on daily dialogues data or any data I guess. I did find this recent paper on fine-tuning DialoGPT with promotes which is supposed to generate "more novel and diverse responses with a higher BLEU score while training only a small portion of the DialoGPT/GPT-2 parameters". So this could save on training time. I'd be keen to work on this together

[racheledelstein]

That sounds good

[RyanSelesnik]

Another potentially useful thread here

[racheledelstein]

And here which they linked in the DialoGPT repo

[RyanSelesnik]

Oh cool yeah I've seen that. Though from what I've been reading now fine-tuning actually doesn't solve the problem since the "fine-tuning process robs [GPT2] of the ability to produce diverse responses, eventually reducing them to dull conversation partners.