Scaling Sentence Embeddings with Large Language Models

Overview

Large language models (LLMs) have recently garnered significant interest. With in-context learning, LLMs achieve impressive results in various natural language tasks. However, the application of LLMs to sentence embeddings remains an area of ongoing research. In this work, we propose an in-context learning-based method aimed at improving sentence embeddings performance. Our approach involves adapting the previous prompt-based representation method for autore- gressive models, constructing a demonstration set that enables LLMs to perform in-context learning, and scaling up the LLMs to different model sizes. Through extensive experiments, in-context learning enables LLMs to generate high-quality sentence embeddings without any fine-tuning. It helps LLMs achieve performance comparable to current contrastive learning methods. By scaling model size, we find scaling to more than tens of billion parameters harms the performance on semantic textual similarity (STS) tasks. However, the largest model outperforms other counterparts and achieves the new state-of-the-art result on transfer tasks. We also fine-tune LLMs with current contrastive learning approach, and the 2.7B OPT model, incorporating our prompt-based method, surpasses the performance of 4.8B ST5, achieving the new state-of-the-art results on STS tasks.

Results on STS Tasks with in-context learning (without fine-tuning)

To evaluate the above results, please run the following script,

bash run_icl.sh [opt-125m|opt-350m|opt-1.3b|opt-2.7b|opt-6.7b|opt-13b|opt-30b|opt-66b]

Results on STS Tasks with contrastive learning (with fine-tuning)

To evaluate the above results, please run the following script:

MODEL_PATH=facebook/opt-2.7b # or  decapoda-research/llama-x-hf  x model size 7b 13b 
LORA=royokong/prompteol-opt-2.7b # or royokong/prompteol-llama-x x model size 7b 13b
TEMPLATE='This_sentence_:_"*sent_0*"_means_in_one_word:"'
python evaluation.py \
    --model_name_or_path   $MODEL_PATH \
    --mode test --mask_embedding_sentence \
    --mask_embedding_sentence_template $TEMPLATE --lora_weight $LORA --load_kbit 16 

Examples

1. Loading base model

   
   import torch
   from transformers import AutoModelForCausalLM, AutoTokenizer

Import our models. The package will take care of downloading the models automatically

tokenizer = AutoTokenizer.from_pretrained("facebook/opt-2.7b") model = AutoModelForCausalLM.from_pretrained("facebook/opt-2.7b") tokenizer.pad_token_id = 0 tokenizer.padding_side = "left" texts = [ "There's a kid on a skateboard.", "A kid is skateboarding.", "A kid is inside the house." ]

### Use in-context learning to generate embeddings
Directly using in-contex learning get embeddings
``` python
template = 'This_sentence_:_"A_jockey_riding_a_horse."_means_in_one_word:"Equestrian".This_sentence_:_"*sent_0*"_means_in_one_word:"'
inputs = tokenizer([template.replace('*sent_0*', i).replace('_', ' ') for i in texts], padding=True,  return_tensors="pt")
with torch.no_grad():
    embeddings = model(**inputs, output_hidden_states=True, return_dict=True).hidden_states[-1][:, -1, :]

Use contrastive learning models to generate embeddings

Using trained LoRA to get embeddings

from peft import PeftModel
peft_model = PeftModel.from_pretrained(model, "royokong/prompteol-opt-2.7b", torch_dtype=torch.float16)
template = 'This_sentence_:_"*sent_0*"_means_in_one_word:"'
inputs = tokenizer([template.replace('*sent_0*', i).replace('_', ' ') for i in texts], padding=True, return_tensors="pt")
with torch.no_grad():
    embeddings = peft_model(**inputs, output_hidden_states=True, return_dict=True).hidden_states[-1][:, -1, :]

Setup

Install Dependencies

pip install -r requirements.txt

Download Data

cd SentEval/data/downstream/
bash download_dataset.sh
cd -
cd ./data
bash download_nli.sh
cd -

In-context learning

We provide in-context learning examples in icl_examples.txt.

To evaluate examples on STS-B development set

BASE_MODEL=facebook/opt-2.7b
python evaluation.py \
   --model_name_or_path $BASE_MODEL \
   --mode dev --mask_embedding_sentence \
   --load_kbit 4 --icl_examples_file 274_templates.txt

Contrastive learning

Train

bash train_llm.sh opt-2.7b # can be other models

Test

bash eval_checkpoints.sh opt-2.7b-lora # first evaluate checkpoint on STS-B dev. and evaluate best checkpoint on STS tasks

Acknowledgement

Our Code is based on SimCSE and alpaca-lora