DedalusHealthCare/mlm_demo_mu

Demo MLM Dataset (MU)

Dataset Description

This dataset contains clinical texts prepared for Masked Language Modeling (MLM) training in Multilingual language. The dataset is designed for pre-training or fine-tuning language models on domain-specific medical text.

Dataset Statistics

• Total examples: 1,200
• Training examples: 1,000
• Validation examples: 100
• Test examples: 100
• Average text length: 0.0 characters
• Max text length: 155 characters
• Min text length: 65 characters

Language Distribution

• en: 1,200 examples

Dataset Structure

Each example contains:

• text: The input text for MLM training
• meta: Optional metadata (e.g., language information)

Usage

from datasets import load_dataset

dataset = load_dataset("DedalusHealthCare/mlm_demo_mu")

# Access the data
train_data = dataset["train"]
print(train_data[0])  # Show first example

# Example output format:
# {
#     'text': 'Patient presents with symptoms of diabetes and hypertension.',
#     'meta': {'lang': 'en'}
# }

Training

This dataset is suitable for:

• Masked Language Modeling (MLM) pre-training
• Domain adaptation of existing language models to clinical text
• Fine-tuning BERT-like models on medical vocabulary

Example MLM Training

from transformers import AutoTokenizer, AutoModelForMaskedLM, TrainingArguments, Trainer
from transformers import DataCollatorForLanguageModeling

# Load model and tokenizer
model_name = "bert-base-uncased"  # or your preferred base model
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForMaskedLM.from_pretrained(model_name)

# Tokenize the dataset
def tokenize_function(examples):
    return tokenizer(examples["text"], truncation=True, padding=True, max_length=512)

tokenized_dataset = dataset.map(tokenize_function, batched=True)

# Data collator for MLM
data_collator = DataCollatorForLanguageModeling(
    tokenizer=tokenizer, mlm=True, mlm_probability=0.15
)

# Training arguments and trainer setup
training_args = TrainingArguments(
    output_dir="./mlm-model",
    overwrite_output_dir=True,
    num_train_epochs=3,
    per_device_train_batch_size=16,
    save_steps=10_000,
    save_total_limit=2,
)

trainer = Trainer(
    model=model,
    args=training_args,
    data_collator=data_collator,
    train_dataset=tokenized_dataset["train"],
    eval_dataset=tokenized_dataset["validation"],
)

# Train the model
trainer.train()

Citation

If you use this dataset, please cite:

@dataset{demo_mu_mlm,
  title = {Demo MLM Dataset (MU)},
  author = {DH Healthcare GmbH},
  year = {2025},
  publisher = {Hugging Face},
}

License

This dataset is proprietary and owned by DH Healthcare GmbH. All rights reserved.