We evaluate the quality of nomic-bert-2048 on the standard GLUE benchmark. We find it performs comparably to other BERT models but with the advantage of a significantly longer context length.
We use BookCorpus and a 2023 dump of wikipedia. We pack and tokenize the sequences to 2048 tokens. If a document is shorter than 2048 tokens, we append another document until it fits 2048 tokens. If a document is greater than 2048 tokens, we split it across multiple documents. We release the dataset here
from transformers import AutoModelForMaskedLM, AutoConfig, AutoTokenizer, pipeline
tokenizer = AutoTokenizer.from_pretrained('bert-base-uncased') # `nomic-bert-2048` uses the standard BERT tokenizer
config = AutoConfig.from_pretrained('nomic-ai/nomic-bert-2048', trust_remote_code=True) # the config needs to be passed in
model = AutoModelForMaskedLM.from_pretrained('nomic-ai/nomic-bert-2048',config=config, trust_remote_code=True)
# To use this model directly for masked language modeling
classifier = pipeline('fill-mask', model=model, tokenizer=tokenizer,device="cpu")
print(classifier("I [MASK] to the store yesterday."))
To finetune the model for a Sequence Classification task, you can use the following snippet
from transformers import AutoConfig, AutoModelForSequenceClassification
model_path = "nomic-ai/nomic-bert-2048"
config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
# strict needs to be false here since we're initializing some new params
model = AutoModelForSequenceClassification.from_pretrained(model_path, config=config, trust_remote_code=True, strict=False)
unsloth 
mlx-community