NanoBERT-V2

This model was pre-trained from scratch using the same model architecture as google-bert/bert-base-uncased.

A custom tokenizer was used, created using wordpiece from a corpus consisting of 200,000 domain-specific papers, and a cut of BookCorpus.

Intended uses & limitations

Intended for training on downstream tasks using Nanoscience datasets. Can be used directly to create dense vector representations for information retrieval.

Training and evaluation data

Trained using 2 nodes on Polaris: https://docs.alcf.anl.gov/polaris/hardware-overview/machine-overview/

Example usage

MLM

from transformers import pipeline

mlm_pipeline = pipeline("fill-mask", model="Flamenco43/NanoBERT-V2")

text = "The most common nanoparticle is [MASK]."

predictions = mlm_pipeline(text)

for prediction in predictions: print(f"Token: {prediction['token_str']}, Score: {prediction['score']:.4f}")

Embedding generation (using the HF api)

api_token = "your_api_token"

model_name = "Flamenco43/NanoBERT-V2"

url = f"https://api-inference.huggingface.co/pipeline/feature-extraction/{model_name}"

headers = {"Authorization": f"Bearer {api_token}"}

data = { "inputs": "This is a sample text for embedding generation." }

response = requests.post(url, headers=headers, json=data)

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 32
• eval_batch_size: 32
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• total_train_batch_size: 256
• total_eval_batch_size: 256
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• training_steps: 1000000

Framework versions

• Transformers 4.41.2
• Pytorch 2.3.0
• Datasets 2.20.0
• Tokenizers 0.19.1