CURP: Codebook-based Continuous User Representation for Personalized Generation with LLMs

Introduction

This repository contains the implementation of CURP (Codebook-based Continuous User Representation for Personalized Generation with LLMs), a novel framework for learning interpretable and continuous user representations to enhance personalized text generation with large language models (LLMs). The training process can be mainly devided into 2 stages. In the first stage we construct a universal codebook via product quantization with balanced K-Means initialization to build user prototype. In the second stage we project the prototype representation into LLM's space and guide for personalization.

Installation

git clone https://github.com/RaidonWong/CURP.git curp
cd curp
pip install -r requirements.txt

Data Preprocessing

1. AlignX Dataset

• Source: AlignX
• Fields Used: "prompt", "chosen", "rejected", "Demographic Information", "User-Generated Content"
• Note: We use a randomly filtered deduplicated subset. Can be replaced with any user history dataset with broad knowledge distribution.

2. Val Datasets

• Sources:
  • LaMP Benchmark
  • LongLaMP Benchmark
• Tasks: News Headline, Tweet Paraphrase, Review Writing
• Preprocessing:
  • News Headline: Filter short texts; use LLaMA-3-8B to judge headline-paragraph association.
  • Tweet Paraphrase: Remove noisy entries (e.g., @<ref>); use LLM for filter.
  • Review Writing: Keep consistent-rating reviews, length filter

Usage

• First of all, you need to download the Qwen-2.5-7B-Instruct model and Contriever model
  • Qwen-2.5-7B-Instruct
  • Contriever Then you need to add two speicial token, <PAD> and <USR_EMB> and resize the LLM tokenizer embedding input. The <PAD> is for padding and the <USR_EMB> is for indicating the place to insert our user model.

special_tokens = {
    "additional_special_tokens": ["<PAD>", "<USR_EMB>"]
}
tokenizer.add_special_tokens(special_tokens)
model.resize_token_embeddings(len(tokenizer))

-Secondly, you need to prepare the embedding of each history for next steps by:

python encode.py

• Thirdly, you need to pretrain a Product Quantized codebook by

bash step1.sh

• Then, you need to align the representation with LLMs by

bash step2.sh

• After training, you can inference by

python inference.py

For other tasks, they share a similar process. We also provide corresponding code.

The argument is by default. If you want to change, you can change as you need.