activeslr-picos-ner

A large-scale synthetic dataset for biomedical Named Entity Recognition (NER) of PICOS elements in clinical research abstracts.

Dataset Description

This dataset contains 21,010 PubMed abstracts annotated with five PICOS entity types commonly used in evidence-based medicine and systematic reviews.

PICOS Elements

Entity	Description	Example
Population	Patient demographics, conditions, sample size	"99 patients with type 2 diabetes"
Intervention	Treatments, drugs, procedures, therapies	"metformin 500mg twice daily"
Comparator	Control groups, alternative treatments	"placebo group"
Outcome	Measured results, endpoints, effects	"HbA1c reduction of 1.2%"
Study Design	Research methodology type	"randomized controlled trial"

Dataset Statistics

Split	Documents
Train	18,909
Test	2,101
Total	21,010

Usage

Load Dataset

from datasets import load_dataset

dataset = load_dataset("anupshaw/activeslr-picos-ner")

# Access splits
train_data = dataset["train"]
test_data = dataset["test"]

print(f"Train examples: {len(train_data)}")
print(f"Test examples: {len(test_data)}")

Example Record

example = dataset["train"][0]

print(f"PMID: {example['pmid']}")
print(f"Tokens: {example['tokens'][:10]}")
print(f"Population: {example['population'][:10]}")
print(f"Intervention: {example['intervention'][:10]}")

Data Schema

Each record contains:

Field	Type	Description
pmid	string	PubMed ID
text	string	Raw abstract text
tokens	List[string]	Tokenized abstract text
pos_tags	List[string]	Penn Treebank POS tags
population	List[int]	Binary labels (0/1)
intervention	List[int]	Binary labels (0/1)
comparator	List[int]	Binary labels (0/1)
outcome	List[int]	Binary labels (0/1)
study_design	List[int]	Binary labels (0/1)

Label Format

• 0 = Token is NOT part of the entity
• 1 = Token IS part of the entity

Note: Labels are not mutually exclusive - a token can belong to multiple entity types simultaneously.

Training Example

from datasets import load_dataset
from transformers import (
    AutoTokenizer,
    AutoModelForTokenClassification,
    TrainingArguments,
    Trainer
)

# Load dataset
dataset = load_dataset("anupshaw/activeslr-picos-ner")

# Load model
model_name = "microsoft/BiomedNLP-BiomedBERT-base-uncased-abstract"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForTokenClassification.from_pretrained(
    model_name,
    num_labels=11  # O + 5 entity types x 2 (B/I tags)
)

# See full training script in the repository

Dataset Generation

This dataset was generated using:

• Source: PubMed abstracts from uiyunkim-hub/pubmed-abstract
• Annotation: LLM-based extraction (gpt-oss-120b via Baseten)
• Tokenization: EBM-NLP compatible tokenization
• POS Tagging: NLTK with Penn Treebank tagset

Generation Pipeline

PubMed Abstracts → LLM Annotation → Tokenization → POS Tagging → Validation

Related Datasets

• EBM-NLP - Human-annotated PICO dataset (4,993 abstracts)
• PICO-Corpus - PICO detection dataset

Citation

@dataset{activeslr_picos_ner_2026,
  title = {activeslr-picos-ner: A Synthetic PICOS NER Dataset},
  author = {Anup Shaw},
  year = {2026},
  publisher = {HuggingFace Hub},
  url = {https://huggingface.co/datasets/anupshaw/activeslr-picos-ner}
}

License

This dataset is released under the Apache 2.0 License.

• Source abstracts are from PubMed (public domain)
• Annotations are synthetically generated

Contact

For questions or issues, please open an issue on GitHub.