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)

Model	STS12	STS13	STS14	STS15	STS16	STSb	SICK-R	Avg.
OPT 125M	62.22	73.10	61.84	71.09	72.08	67.80	64.10	67.46
OPT 350M	63.87	73.85	63.41	72.45	73.13	70.84	65.61	69.02
OPT 1.3B	72.78	83.77	73.61	83.42	80.60	78.80	69.69	77.52
OPT 2.7B	68.49	84.72	75.15	83.62	81.34	80.94	72.97	78.18
OPT 6.7B	70.65	84.51	75.01	83.51	82.00	81.12	76.77	79.08
OPT 13B	71.99	85.22	76.04	82.23	81.38	81.42	75.00	79.04
OPT 30B	69.99	83.35	74.75	83.14	82.42	81.45	77.46	78.94
OPT 66B	69.93	83.29	74.88	80.10	81.11	81.76	76.26	78.19

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)

Model	STS12	STS13	STS14	STS15	STS16	STSb	SICK-R	Avg.
royokong/prompteol-opt-1.3b	79.01	89.26	84.10	88.30	84.62	87.71	80.52	84.79
royokong/prompteol-opt-2.7b	79.49	89.64	84.80	89.51	85.91	88.33	81.64	85.62
royokong/prompteol-opt-6.7b	80.14	90.02	84.94	89.78	85.84	88.75	81.29	85.82
royokong/prompteol-opt-13b	80.20	90.24	85.34	89.52	85.90	88.56	82.06	85.97

royokong/prompteol-llama-7b	79.16	90.22	85.40	88.99	86.25	88.37	81.51	85.70
royokong/prompteol-llama-13b	78.63	90.03	85.46	89.48	86.18	88.45	82.69	85.85

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

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

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

hppRC/scaling_sentemb