Obfuscated clinical documents and derived extraction prompts. Access is granted manually. By requesting access you confirm you will use it only for authorised research, will not attempt re-identification, and will not redistribute it.

Oncology extraction — Dedalus ontology

Structured extraction from European oncology documents (mostly German and French pathology reports) against a runtime-supplied field specification: the ontology arrives in the prompt, so one model must obey whichever spec it is given.

split	rows	`output` (training target)	`reference` (answer key)
`train`	2,287	2,287 labelled	—
`eval`	103	empty by design	57 rows / 30% of field-slots

instruction + "\n\n---\n" + input reconstructs the original prompt exactly.

Columns

output and reference are nested lists of structs, not JSON strings, so the dataset viewer expands them. Each element:

key	type
`field_name`	string	the ontology field
`category`	string	its ontology category
`raw_value`	string	as the document states it, in the source language
`normalized_value`	string	standards-conformant, English, plus code where one applies
`explanation`	string	English reasoning for the value
`reasoning_excerpt`	string	verbatim quote, in the document's own language
`confidence_score`	double	0.0-1.0

A field's value may legitimately be a string, a number or a list. A typed column cannot be all three, so raw_value / normalized_value are strings and list values are JSON-encoded inside them ('["BRCA1", "TP53"]'). null is preserved as null, and means the document does not address the field.

Also present: ontology, corpus, strategy, doc_type, language, requested_fields, source_file, source_sha256.

`corpus` is not always `dedalus` — read this before filtering

Every row uses the Dedalus field specification, but 28.9% of train rows (662) and 9 eval rows are MSK source documents, all with strategy = cross_ontology. That pairing is deliberate: it trains the model to obey whichever ontology arrives in the prompt rather than memorising one corpus. The split is keyed on the ontology the prompt was rendered from, not the document's origin.

For Dedalus source documents only:

ds = ds.filter(lambda r: r["corpus"] == "dedalus")   # train 1,625 · eval 94

The eval split: partial reference, and read this before using it

eval is a frozen holdout (from 160 documents: 100 Dedalus + 60 MSK; 103 render as Dedalus-ontology rows). Train/eval document overlap is 0, verified by sha256.

output is empty on this split by design — it is a holdout, not a training target. A separate reference column carries whatever golden answers exist:


rows with any reference	57 of 103 (55%)
field-slots requested	1,748
field-slots with a reference	522 (30%)

Two per-row integers, reference_n and reference_of_requested, make the coverage explicit so nothing has to be inferred from an empty string.

Why coverage is partial. 46 documents were carved out before any teacher call and were never labelled — there is genuinely no key for them. The other 57 entered the holdout later, having previously been in the training pool, so a label exists. Even for those, the eval prompt samples a different field subset than the original render did, so only the requested fields that happen to be covered are carried over.

Do not use reference to score a model that was trained on this corpus. Those 57 documents were in earlier training sets; for such a model the reference measures recall of seen items. It is a clean reference only for a model trained on data that excludes this holdout.

Label repairs applied

836 defects were located, verified against the source documents, and fixed:

defect	cells
numeric values stored as text (`"3 cm"`, `"1.2 mm"`) where a number is required	299
`pr_status` citing evidence that never mentions progesterone	143
`stage_group` holding an Ann Arbor / Lugano stage	56
stage codes (`pT3`, `cT4b`) in a free-text anatomy field	44
melanoma thickness filed as tumour diameter	36
`tumor_size` read off a T-category, or from an endoscopic distance	10
`mitotic_rate` from an Elston-Ellis sub-score	7
`isup_grade_group` misplacement, and one Gleason 3+4 mapped to GG3	8
`closest_margin_distance` from a planned excision width	6

Found by deterministic extractors that read the source documents directly and were validated against the labels first — 100% agreement on the five measurement fields, 93.8–98.4% on grading/staging.

Known remaining defects — not fixed

47 histologic_grade labels are not normalised to G-notation ("poorly differentiated", "Fuhrman nuclear grade 3" left raw).
25 staging_system_edition labels are inconsistent on the compound token UICC/AJCC 8. Auflage — 176 read as UICC, 8 as AJCC, from identical text.

Both need a normalisation decision rather than a deletion, so they were left alone rather than guessed at.

Caveats

Labels are teacher-generated (deepseek-v4-flash, plus verification, merging and repair passes). They are not independently adjudicated ground truth. An independent check found two systems agreeing on 92.9% of blank cells in both directions, but agreement is not correctness.
The corpus leaves roughly half of requested cells null. Much of that is correct abstention — the source document genuinely does not address the field — but the share that is a genuine miss has not been measured.
Dedalus documents are synthetic/templated and ~95% German. Codes are masked to [CODE], so coding fields have no quotable in-document evidence.

Downloads last month: -

Total file size:

4.68 MB

You need to agree to share your contact information to access this dataset