bert-base-1024-biencoder-6M-pairs

A long context biencoder based on MosaicML's BERT pretrained on 1024 sequence length. This model maps sentences & paragraphs to a 768 dimensional dense vector space and can be used for tasks like clustering or semantic search.

Usage

Download the model and related scripts

git clone https://huggingface.co./shreyansh26/bert-base-1024-biencoder-6M-pairs

Inference

import torch
from torch import nn
from transformers import AutoModelForMaskedLM, AutoTokenizer, pipeline, AutoModel
from mosaic_bert import BertModel

# pip install triton==2.0.0.dev20221202 --no-deps if using Pytorch 2.0

class AutoModelForSentenceEmbedding(nn.Module):
    def __init__(self, model, tokenizer, normalize=True):
        super(AutoModelForSentenceEmbedding, self).__init__()

        self.model = model.to("cuda")
        self.normalize = normalize
        self.tokenizer = tokenizer

    def forward(self, **kwargs):
        model_output = self.model(**kwargs)
        embeddings = self.mean_pooling(model_output, kwargs['attention_mask'])
        if self.normalize:
            embeddings = torch.nn.functional.normalize(embeddings, p=2, dim=1)

        return embeddings

    def mean_pooling(self, model_output, attention_mask):
        token_embeddings = model_output[0]  # First element of model_output contains all token embeddings
        input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
        return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)

model = AutoModel.from_pretrained("<path-to-model>", trust_remote_code=True).to("cuda")
model = AutoModelForSentenceEmbedding(model, tokenizer)
tokenizer = AutoTokenizer.from_pretrained('bert-base-uncased')

sentences = ["This is an example sentence", "Each sentence is converted"]

encoded_input = tokenizer(sentences, padding=True, truncation=True, max_length=1024, return_tensors='pt').to("cuda")
embeddings = model(**encoded_input)

print(embeddings)
print(embeddings.shape)

Other details

Training

This model has been trained on 6.4M randomly sampled pairs of sentences/paragraphs from the same training set that Sentence Transformers models use. Details of the training set here.

The training (along with hyperparameters), inference and data loading scripts can all be found in this Github repository.

Evaluations

We ran the model on a few retrieval based benchmarks (CQADupstackEnglishRetrieval, DBPedia, MSMARCO, QuoraRetrieval) and the results are here.