search

tonychenxyz / selfie

This repository contains the code and data for the paper "SelfIE: Self-Interpretation of Large Language Model Embeddings" by Haozhe Chen, Carl Vondrick, and Chengzhi Mao.
24 stars 2 forks source link

Project structure #1

Open dar-tau opened 3 months ago

dar-tau commented 3 months ago

Hi guys, great work!! Quick question about the code structure, what is the purpose of generate_wrappers.py? (couldn't understand what's the diff from the original) What was added in llama_forward_wrappers? I ask because I want to work on something similar, but I will have to work on other models.

Thanks :)

tonychenxyz commented 3 months ago

Hi! Thanks!

The main purpose of generate_wrappers.py is to introduce generate_interpret, which is same as model.generate except it accepts embedding vectors and instructions about where to replace them in forward pass. Calling generate_interpret(model, model_inputs, ..., insert_info) carries out the usual generation pipeline except it conducts the embedding replacement at forward pass according to instructions in insert_info, which make it generate interpretation of the inserted embeddings. All functions that model.generate calls and requires inserting embedding are copied into this file and modified accordingly.

llama_forward_wrappers does similar thing to model.forward (wrapped into model_forward_interpret).

Let me know if I can explain more.