dataset_info:
features:
- name: sentence1
dtype: string
- name: labels
dtype: string
- name: metadata
dtype: string
splits:
- name: train
num_bytes: 578657578
num_examples: 2211861
- name: validation
num_bytes: 7601235
num_examples: 28932
- name: test
num_bytes: 7683141
num_examples: 29493
download_size: 224972141
dataset_size: 593941954
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: validation
path: data/validation-*
- split: test
path: data/test-*
license: other
source_datasets:
- zj88zj/PubMed_200k_RCT
pubmed-rct
PubMed 200k RCT sentence roles, retaining native PMID/abstract partitions.
Converts the pinned plaintext training Hub copy, preserving PMID headers and numbers. Uses original dev/test files because the Hub dev file is RTF, not plaintext. Streaming conversion avoids collecting 2.2 million sentence dictionaries in memory. Native train/dev/test are preserved and verified PMID-disjoint. Original data distribution: https://github.com/Franck-Dernoncourt/pubmed-rct. Terms remain unspecified: the original README warns that PubMed abstracts can be copyrighted; free availability does not establish redistribution rights.
Original data: zj88zj/PubMed_200k_RCT, original distribution. Repackaged as parquet for tasksource by scripts/repackage_dataset/specialized_text.py.
Conversion provenance: source revisions, hashes, configuration and exclusions.