README.md

metadata

license: llama2

RepLLaMA-7B-Passage-MRL

Fine-Tuning LLaMA for Multi-Stage Text Retrieval. Xueguang Ma, Liang Wang, Nan Yang, Furu Wei, Jimmy Lin, arXiv 2023

This model is fine-tuned from LLaMA-2-7B using LoRA and the embedding size is flexible, as Matryoshka Representation Learning is applied during training. The maximum dimensionality of query and passage embedding is 4096.

Training Data

The model is fine-tuned on the training split of MS MARCO Passage Ranking datasets for 1 epoch. Please check our paper for details.

Usage

Below is an example to encode a query and a passage, and then compute their similarity using their embedding.

import torch
from transformers import AutoModel, AutoTokenizer

# Load the tokenizer and model
tokenizer = AutoTokenizer.from_pretrained('meta-llama/Llama-2-7b-hf')
model = AutoModel.from_pretrained('castorini/repllama-v1-mrl-7b-lora-passage')
dim = 512

# Define query and passage inputs
query = "What is llama?"
title = "Llama"
passage = "The llama is a domesticated South American camelid, widely used as a meat and pack animal by Andean cultures since the pre-Columbian era."
query_input = tokenizer(f'query: {query}</s>', return_tensors='pt')
passage_input = tokenizer(f'passage: {title} {passage}</s>', return_tensors='pt')

# Run the model forward to compute embeddings and query-passage similarity score
with torch.no_grad():
    # compute query embedding
    query_outputs = model(**query_input)
    query_embedding = query_outputs.last_hidden_state[0][-1][:dim]
    query_embedding = torch.nn.functional.normalize(query_embedding, p=2, dim=0)

    # compute passage embedding
    passage_outputs = model(**passage_input)
    passage_embeddings = passage_outputs.last_hidden_state[0][-1][:dim]
    passage_embeddings = torch.nn.functional.normalize(passage_embeddings, p=2, dim=0)

    # compute similarity score
    score = torch.dot(query_embedding, passage_embeddings)
    print(score)

Batch inference and training

An unofficial replication of the inference and training code can be found here

Citation

If you find our paper or models helpful, please consider cite as follows:

@article{rankllama,
      title={Fine-Tuning LLaMA for Multi-Stage Text Retrieval}, 
      author={Xueguang Ma and Liang Wang and Nan Yang and Furu Wei and Jimmy Lin},
      year={2023},
      journal={arXiv:2310.08319},
}