BAAI

/

bge-m3

like 3.2k

Follow

Beijing Academy of Artificial Intelligence 3.95k

Hi there! I'm testing your model in a spanish RAG application and it is for far the best spanish model i have ever tried, using your model to retrieve and rerank works like a charm! very nice JOB!

in my tests I have find some extrange behaviours:

1. the embeddings results values changes a little bit when I switch from batchsize = 1 to > 1
2. when I have a batchsize > 1 and I call first the compute_score works fine, but when I call encode it hangs the process in a deadlock. (only happens in that secuence order of calls)
3. I debug the model code and I find that the Dataloader class in the encode method has the deadlock problem, I do not known too much of pytorch so I looked to the for iterator in the compute_score and that method do not use a DataLoader, instead use direct the tokenizer. I adapted de code of encode (based on the compute_score method):

from:

all_dense_embeddings, all_lexical_weights, all_colbert_vec = [], [], []
    for batch_data in tqdm(data_loader, desc='encoding', mininterval=10):
        batch_data = batch_data.to(self.device)

        output = self.model(batch_data,
                            return_dense=return_dense,
                            return_sparse=return_sparse,
                            return_colbert=return_colbert_vecs)

to:

all_dense_embeddings, all_lexical_weights, all_colbert_vec = [], [], []
    for start_index in tqdm(range(0, len(sentences), batch_size), desc="Encoding", disable=len(sentences) < 0):
        sentences_batch = sentences[start_index:start_index + batch_size]
        queries_inputs = self.tokenizer(sentences_batch,
                   max_length=max_length,
                   padding=True,
                   return_token_type_ids=False,
                   truncation=True,
                   return_tensors='pt').to(self.device)
        output = self.model(queries_inputs,
                            return_dense=return_dense,
                            return_sparse=return_sparse,
                            return_colbert=return_colbert_vecs)

And the problem of the deadlock gone.

Hope my findings help you to improve your great JOB!

Best regards