At 10:00AM PT: Author of ReFT: Representation Finetuning, at Oxen.ai Paper Club

3 points

2 years ago

Arxiv paper first author Zhengxuan Wu will join Greg Schoeninger on today's 10:00 AM Pacific Time Oxen.AI Paper Club. Topic: How editing representations can be better than Parameter-efficient finetuning (PEFT) methods. https://lu.ma/oxen

ReFT: Representation Finetuning for Language Models.

Greg, 3 questions, 1 comment from reading just the abstract.

What exactly is meant by a "representation"? i.e., what part of the neural network is captured by what this paper refers to as a representation?

What is meant by an intervention in "task specific intervention" ? I haven't heard that term before with pretraining or fine-tuning.

In API terms, would the point of this paper be like saying: "Instead of improving a RESTful API by improving the fidelity (depth) or breadth of sniff-able inputs and outputs, we will improve the API by directly changing the guts of the code for all existing inputs and outputs?"

Comment) The abstract makes this paper sound very voodoo. Hope testing was apples :: apples.

Look forward to your demystification on Friday, Greg. So cool that you have the paper's first author joining you to help explain and answer questions.

Deets:

https://lu.ma/oxen

Friday July 19, 10:00 AM Pacific, 1:00 PM Eastern Time on Zoom

Paper: https://arxiv.org/pdf/2404.03592

Gratitude: Thank you Greg Schoeninger, Scott Howard and the Oxen team for giving me an Easy button. You share your knowledge with the community while providing cool tools to curate datasets at oxen.ai.