Skip to content

LCRec

Adapting Large Language Models by Integrating Collaborative Semantics for Recommendation.

Overview

LCRec adapts large language models (LLMs) for recommendation by integrating collaborative semantics through codebook tokens. It uses semantic IDs from RQ-VAE to represent items and fine-tunes an LLM to generate item recommendations.

Architecture

Item Text → RQ-VAE → Semantic IDs [<C0_x>, <C1_y>, ..., <C4_z>]
                          ↓
User History + Prompt → LLM (Qwen) → Generated Semantic IDs
                          ↓
                    Constrained Decoding
                          ↓
                    Recommended Items

Key Components

  • Semantic ID Generation: RQ-VAE with 5 codebooks (256 codes each)
  • Codebook Tokens: Special tokens <Ci_j> added to LLM vocabulary
  • Constrained Decoding: Beam search with prefix constraints
  • Multi-task Training: seqrec, item2index, index2item tasks

Training Pipeline

Step 1: Train RQ-VAE for Semantic IDs

python genrec/trainers/rqvae_trainer.py config/lcrec/amazon/rqvae.gin --split beauty

Step 2: Fine-tune LLM

python genrec/trainers/lcrec_trainer.py config/lcrec/amazon/lcrec.gin --split beauty

Debug Mode (Quick Testing)

python genrec/trainers/lcrec_trainer.py config/lcrec/amazon/lcrec_debug.gin

Configuration

# config/lcrec/amazon/lcrec.gin

# Training
train.epochs = 4
train.batch_size = 32
train.learning_rate = 2e-5
train.max_length = 512

# Model
train.pretrained_path = %MODEL_HUB_QWEN3_1_7B
train.use_lora = False

# Codebook
train.num_codebooks = 5
train.codebook_size = 256

# Evaluation
train.eval_beam_width = 10

Tasks

LCRec supports three training tasks:

Task Input Output
seqrec User history Next item semantic ID
item2index Item text Item semantic ID
index2item Semantic ID Item text

Model API

from genrec.models import LCRec

model = LCRec(pretrained_path="Qwen/Qwen2.5-1.5B")
model.add_codebook_tokens(num_codebooks=5, codebook_size=256)

# Generate recommendations
outputs = model.generate(
    input_ids=prompt_ids,
    max_new_tokens=6,
    num_beams=10,
)

Reference