I recall Jeremy doing an interesting demo where he created a new document from just a few keywords by using a pre-trained language model fine tuned with a large corpus of academic document abstracts. I can’t find it in the course material. Can someone please let me know which lesson that was in and, ideally, if there was a demo notebook? I’ve looked at all the 2018 course #1 and #2 notebooks and can’t find it.
The demo results looked something like what might be produced by OpenAI’s gpt-2. I want to fine tune a ULMFit model with my own corpus.
You can find it here in this notebook.
Thanks Seemant. Looks like that notebook deal exclusively with the iMDB dataset but maybe I haven’t looked closely enough. Also possible that Jeremy did the academic paper demo without sharing the notebook. I’ll keep digging.
.predict will work with any language model that you generate or load. You give it a word or set of words, it will then predict the next one and continue to do that till you tell it to stop (number of words.) This notebook above is a single example of that.
You can also find an example of using a beam search to predict the next work. All in the docs. This should all work “out of the box” with the Wiki pre-trained model and you could fine-tune to your particular needs.
Great info Bobak. Thank you very much!