许可证: mit 数据集:

• KomeijiForce/Inbedder-Pretrain-Data 语言:
• en

[ACL2024] 答案即所需：通过回答问题实现指令跟随的文本嵌入

InBedder🛌 是一款专为遵循指令而设计的文本嵌入器。这种能够遵循指令的文本嵌入器可以捕捉用户指令所指定的文本特征。InBedder 提供了一个新颖的视角，将指令视为关于输入文本的问题，并通过编码预期答案来相应地获取表示。我们展示了 InBedder 能够识别不同评估任务中的指令。

以下是一个来自 https://github.com/zhang-yu-wei/InBedder/blob/main/UseCase.ipynb 的使用案例：

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>)