OncoRx-Bench

OncoRx-Bench is a controlled synthetic dataset for research on medication and oncology-regimen information extraction. The reference release contains 2,000 automatically generated records: 1,600 training rows and 400 test rows. Whole source-template groups are assigned to one split, yielding zero source-template overlap while preserving an exact 80/20 split within every subcategory.

The text and task-selected annotations are synthetic. Automated validation establishes structural, referential, offset, evidence, distribution, and split consistency; it does not establish clinical validity. The dataset is not for clinical use, clinical decision-making, prescribing, or patient care.

Links

• Source and reproducibility workflow: https://github.com/shivanna-ornl/oncorx-bench
• Dataset page: https://huggingface.co/datasets/AbhishekShivanna/oncorx-bench
• Knowledge source: https://github.com/shivanna-ornl/unified-oncology-treatment-database

Data files

Split	Path	Rows
train	data/train.jsonl	1,600
test	data/test.jsonl	400

The split union contains exactly ONCORX-0001 through ONCORX-2000, with no duplicate identifiers and no overlap. The canonical full JSONL SHA-256 is 97a6c2a22eacf0d68e1762051fbe57b943f24c032a014eadd64f7e36aecf952f.

from datasets import load_dataset

dataset = load_dataset("AbhishekShivanna/oncorx-bench")
print(dataset["train"][0])

Schema

Field	Type	Description
sample_id	string	Record identifier
clinical_text	string	Synthetic input text
category	string	Major evaluation category
subcategory	string	Fine-grained evaluation category
difficulty	string	Author-assigned scenario difficulty
drug_mentions	list of objects	Drug evidence spans, normalized labels, context, and optional prescribing attributes
regimen_mentions	list of objects	Regimen evidence spans, normalized labels, projected components, and visible cycle/intent metadata
num_drugs	integer	Number of unique normalized drugs in drug_mentions
note_type	string	Scenario-conditioned synthetic note-type metadata

Drug objects include inclusive start_char and exclusive end_char offsets. The evidence_type field distinguishes a literal explicit_surface from a regimen_inference, for which the stored evidence span is the named regimen. Optional fields are omitted when they are not supported by the rendered text.

Reference-release composition

• 3,525 drug annotation objects, including 2,974 explicit-surface objects and 551 regimen-inference objects
• 170 unique normalized drugs and 433 stored evidence surfaces: 328 literal drug surfaces and 105 regimen-evidence spans
• 1,286 drug objects with prescribing information
• 444 regimen objects spanning 176 normalized regimen names
• 1,110 component assignments spanning 109 unique components

Category	Description	Rows
C1	Core medication extraction	600
C2	Prescribing information	450
C3	Regimen semantics	450
C4	Context and safety language	300
C5	Noise and ambiguity	200

Provenance and reproducibility

The generator consumes three checked-in benchmark views derived by a checksum-verifying adapter from a pinned Unified Oncology Treatment Database revision. The source repository contains the materialized inputs, projection audit, provenance manifest, generator, strict validator, deterministic split exporter, tests, and one-command byte-comparison workflow.

This Hugging Face repository intentionally contains only the dataset card and Viewer-compatible split files. Benchmark generation and validation code lives in the linked GitHub repository.

Limitations

• Records are short English template-generated snippets, not real clinical notes, and they do not model full documents or EHR prevalence.
• No independent clinical adjudication was performed. Knowledge grounding and structural validation do not prove the appropriateness of every disease, drug, dose, route, frequency, schedule, or regimen combination.
• Annotations are task-selected rather than guaranteed exhaustive for every medication-like string in every category.
• The provided split is source-template-disjoint but not drug-, entity-, or regimen-disjoint.
• Author-controlled attribute profiles and surface-variation maps are synthetic benchmark controls, not treatment recommendations.

License and redistribution

No license is declared for this dataset release. Public availability does not supersede the terms of the source knowledge resources. Review the data notice and provenance in the linked GitHub repository before redistribution or commercial use.

Citation

Until the paper is published, cite this dataset repository and the exact Hub revision used. Use the linked GitHub repository for the corresponding source revision and reproduction instructions.