OpenAI's Summarizing Books with Human Feedback

Is this something that will be available via an API endpoint?


Update 2:50 p.m. Pacific: A spokesperson told VentureBeat that “OpenAI has no plans to make the book summarization model publicly available or open source.”

I think this is not difficult to implement…it’s straightforward

  1. break it into sections
  2. make summaries of each
  3. make summary of summaries

I have something similar working on my dev box that uses an extractive summarizer to pick out sentences that add up to 1500 tokens, then summarizes using GPT-3.

1 Like

This was done using reinforcement learning and would at the very least take a lot of time for someone to read. That’s what’s so prohibitive about a lot of nlp tasks is that the datasets just aren’t there and it costs quite a bit to pay to create them or a lot of time doing it yourself. I’m aware of booksum but it would be nice not to have to reinvent the wheel.

I’d love this model to be made public or at the least just one layer of it, I’m summarizing short stories (< 27,000 chars) from PGB for input to a controlled generation process right now.