chongpangnasilemak/icd10cm-exam-openended

ICD-10-CM Exam Open-Ended

70 open-ended ICD-10-CM coding items with published answer keys. The model writes the answer; there are no options to choose from.

The questions are third-party coursework of unverified provenance, reproduced verbatim. The copyright holder is unknown and no licence was granted. If you are the rights holder and object, the dataset will be removed.

No certified medical coder or clinician reviewed any item. The answer keys are the coursework's own, not expert re-verification.

Part of the ICD Capability Evaluation collection. Its closest relative is icd10cm-exam-mcq, which asks for recall of the same code set in multiple-choice form — running both separates recognising a correct code from producing one.

Why open-ended rather than multiple choice

Every sibling dataset in this family is multiple choice, which imposes two ceilings that matter for coding specifically:

• A 25% floor. A model that knows nothing still scores a quarter of the items, and elimination strategies work on the rest.
• Sequencing is invisible. Real ICD-10-CM coding requires the principal diagnosis first. An MCQ either offers the correct order or it does not; it cannot show you a model that picks the right codes and orders them wrongly.

This dataset removes both. The answer to Z51.0, C61 is that string, produced unaided, in that order — and in a dry run one frontier model answered C61, Z51.0: both codes correct, sequence inverted. That failure has no expression in an MCQ.

Scoring

Three measures, reported together because the gaps between them are the diagnostic:

measure	what it says
exact	the code sequence equals the key, order included — the strict measure
unordered	the same codes in any order
points	the coursework's own rubric: 1 per correct code, +½ for correct sequencing

unordered − exact is the sequencing-failure rate: the model found the right codes but could not identify the principal diagnosis. Precision is reported beside points because partial credit rewards a model that emits every plausible code, and precision is what exposes it.

grade.extract_codes reads the answer from the model's reply — preferring an explicit Answer: line, otherwise the last line containing codes — so a reasoning trace that mentions many codes before settling is not scored on the ones it rejected.

Categories

category	test	scored by
code_assignment	33	ordered code sequence
written_answer	12	LLM judge against a published reference
main_term	8	exact term
short_answer	6	exact phrase

written_answer items are paragraphs — a correct answer can share almost no wording with the reference, so no string rule can grade them. gpt-5.6-sol compares the response against the source's published reference answer on a 4-point rubric and returns a justification with every score. The judge is not deciding what is true about ICD-10-CM; it is deciding whether a response conveys what the supplied reference says, which is a far narrower task. A model is refused as its own judge.

Structure

field	notes
id	gen- plus 12 hex characters, derived from question and answer
question	the scenario or question
answer_text	the answer as written — for code items, the codes joined in order
answer_codes	the codes as a list, in required order (empty for written answers)
sequencing_graded	whether the source awarded a sequencing mark for this item
n_codes	length of answer_codes
reference_answer	the published model answer (written_answer items only)
marks	marks the source assigned (written_answer items only)
category / chapter	answer shape, and which exercise set it came from

test = 59 items, dev = 11, disjoint. There is no chance baseline — nothing is chosen — and no answer-position balancing, because there are no positions.

Known corrections to the source

The keys are reproduced verbatim with one exception, recorded on the row itself:

• Chapter 18, prolonged first stage of labor. The source key gave Z3A.38 (38 weeks gestation) against a stem stating 36 weeks. Z3A.36 is a valid FY2023 code and the stem is what a model is shown, so the key was corrected to agree with it.

One key that looks wrong is correct and was left alone: subacute monocytic leukemia in remission → C93.Z1. There is no C93.2 in ICD-10-CM; the Alphabetic Index routes Leukemia > monocytic > specified NEC → C93.Z.

Two items were excluded rather than repaired: a True/False question (no meaningful open-ended answer) and sub-parts asking for free-form commentary rather than a codeable answer.

Limitations

• Not for clinical or billing use.
• Small. 59 test items, and only 33 of them scored by exact match.
• Codes are fiscal-year specific. Checked against FY2023.
• The judged items depend on a model. written_answer scores are one model's reading of whether a response matches a reference, and are not reproducible to the digit.
• No expert review of any item or key.