hazyresearch/M2-BERT-32K-Retrieval-Encoder-V1

Monarch Mixer-BERT

The 80M checkpoint for M2-BERT-32k from the paper Benchmarking and Building Long-Context Retrieval Models with LoCo and M2-BERT.

Check out our GitHub for instructions on how to download and fine-tune it!

How to use

You can load this model using Hugging Face AutoModel:

from transformers import AutoModelForMaskedLM, BertConfig
config = BertConfig.from_pretrained("hazyresearch/M2-BERT-32K-Retrieval-Encoder-V1")
model = AutoModelForMaskedLM.from_pretrained("hazyresearch/M2-BERT-32K-Retrieval-Encoder-V1", config=config, trust_remote_code=True)

This model uses the Hugging Face bert-base-uncased tokenizer:

from transformers import BertTokenizer
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')

How to use

This model generates embeddings for retrieval. The embeddings have a dimensionality of 768:

from transformers import AutoTokenizer, AutoModelForMaskedLM, BertConfig

max_seq_length = 32768
testing_string = "Every morning, I make a cup of coffee to start my day."
config = BertConfig.from_pretrained("hazyresearch/M2-BERT-32K-Retrieval-Encoder-V1")
model = AutoModelForMaskedLM.from_pretrained("hazyresearch/M2-BERT-32K-Retrieval-Encoder-V1", config=config, trust_remote_code=True)

tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased", model_max_length=max_seq_length)
input_ids = tokenizer([testing_string], return_tensors="pt", padding="max_length", return_token_type_ids=False, truncation=True, max_length=max_seq_length)

outputs = model(**input_ids)
embeddings = outputs['sentence_embedding']

Remote Code

This model requires trust_remote_code=True to be passed to the from_pretrained method. This is because we use custom PyTorch code (see our GitHub). You should consider passing a revision argument that specifies the exact git commit of the code, for example:

mlm = AutoModelForMaskedLM.from_pretrained(
   "hazyresearch/M2-BERT-32K-Retrieval-Encoder-V1",
   config=config,
   trust_remote_code=True,
)

Configuration

Note use_flash_mm is false by default. Using FlashMM is currently not supported.