opt-2.7b-InBedder

[ACL2024] Answer is All You Need: Instruction-following Text Embedding via Answering the Question

InBedder🛌 is a text embedder that is designed to follow instructions. Instruction-following text embedder can capture characteristics of texts specified by user instructions. InBedder offers a novel viewpoint that treats the instruction as a question about the input text and encodes the expected answers to obtain the representation accordingly. We show that InBedder is aware of instructions with different evaluation tasks.

The following is a use case from https://github.com/zhang-yu-wei/InBedder/blob/main/UseCase.ipynb

import torch
from torch import nn
from torch.nn.functional import gelu, cosine_similarity
from transformers import AutoTokenizer, AutoModel, AutoModelForMaskedLM

import numpy as np

class InBedder():
    
    def __init__(self, path='KomeijiForce/inbedder-roberta-large', device='cuda:0'):
        
        model = AutoModelForMaskedLM.from_pretrained(path)
    
        self.tokenizer = AutoTokenizer.from_pretrained(path)
        self.model = model.roberta
        self.dense = model.lm_head.dense
        self.layer_norm = model.lm_head.layer_norm
        
        self.device = torch.device(device)
        self.model = self.model.to(self.device)
        self.dense = self.dense.to(self.device)
        self.layer_norm = self.layer_norm.to(self.device)
        
        self.vocab = self.tokenizer.get_vocab()
        self.vocab = {self.vocab[key]:key for key in self.vocab}
        
    def encode(self, input_texts, instruction, n_mask):
        
        if type(instruction) == str:
            prompts = [instruction + self.tokenizer.mask_token*n_mask for input_text in input_texts]
        elif type(instruction) == list:
            prompts = [inst + self.tokenizer.mask_token*n_mask for inst in instruction]
    
        inputs = self.tokenizer(input_texts, prompts, padding=True, truncation=True, return_tensors='pt').to(self.device)

        mask = inputs.input_ids.eq(self.tokenizer.mask_token_id)
        
        outputs = self.model(**inputs)

        logits = outputs.last_hidden_state[mask]
        
        logits = self.layer_norm(gelu(self.dense(logits)))
        
        logits = logits.reshape(len(input_texts), n_mask, -1)
        
        logits = logits.mean(1)
            
        logits = (logits - logits.mean(1, keepdim=True)) / logits.std(1, keepdim=True)
        
        return logits

inbedder = InBedder(path='KomeijiForce/inbedder-roberta-large', device='cpu')

texts = ["I love cat!", "I love dog!", "I dislike cat!"]
instruction = "What is the animal mentioned here?"
embeddings = inbedder.encode(texts, instruction, 3)

cosine_similarity(embeddings[:1], embeddings[1:], dim=1)
# tensor([0.9374, 0.9917], grad_fn=<SumBackward1>)

texts = ["I love cat!", "I love dog!", "I dislike cat!"]
instruction = "What is emotion expressed here?"
embeddings = inbedder.encode(texts, instruction, 3)

cosine_similarity(embeddings[:1], embeddings[1:], dim=1)
# tensor([0.9859, 0.8537], grad_fn=<SumBackward1>)