RadCheck-Bench-ES

Benchmark for evaluating LLMs on detecting internal inconsistencies in Spanish radiology reports: laterality, negation and measurement contradictions verifiable from the text alone. 460 synthetic reports (dev 60, test 340, extra 60), half of the clean ones hard negatives. Code and standalone scorer: GitHub crstian19/radcheck-bench-es.

Hub layout: data/{dev,test,extra}.jsonl are radcheck's evaluation format (the radcheck_bench_v1_*.jsonl files mentioned below). The raw config has the readable view: clean report, corrupted report, labels and metadata. report.html is the self-contained web report. Generation scripts live in the radcheck repository, not here.

About this release (v1)

Independent benchmark for detecting internal inconsistencies in Spanish radiology reports.

Why it exists

radcheck is trained on CARES reports with errors inserted by a rule-based injector, and its own test set is built by the same injector. That test shares the injector's patterns, so its scores are optimistic: the model may learn the injector instead of the task.

This benchmark breaks that dependency: different source texts, different writing styles, errors written like real dictation slips, and many hard negatives. Evaluation only: nothing here is used for training.

What is measured

Input: the full text of one report. Output: the list of inconsistencies, as JSON.

Only contradictions that can be verified from the text alone, without clinical knowledge, are scored:

tipo	Definition
lateralidad (laterality)	The same finding appears with different sides (right/left)
negacion (negation)	One part asserts what another part denies
medida (measurement)	The same measurement differs (number or unit)

Out of scope: whether a finding is clinically right, diagnoses, severity, recommendations. A finding missing from the conclusion and dictation typos go to a separate track (extra) that does not count towards the main score.

Report text, labels and explanations are in Spanish, as in the task.

Files

File	Content
specs.jsonl	460 composition specs, generated with seed 13
raw.jsonl	Full items: clean report, corrupted report, labels, metadata
radcheck_bench_v1_dev.jsonl	60 items, radcheck evaluation format
radcheck_bench_v1_test.jsonl	340 items
radcheck_bench_v1_extra.jsonl	60 items of the out-of-scope track
review.jsonl	Items rejected by the independent check, with the reason
stats.json	Counts per slice, style, modality, region, difficulty and length
radiologist_review.md	50 stratified test items with checkboxes, for the radiologist (Spanish)
report.html	Self-contained web report: composition, acceptance criteria, all 460 items and the rejections
generate.py validate.py build.py check.py report.py	Pipeline
run_generate.sh run_check.sh	Launchers that resolve the API key at run time
prompts/	Every prompt template and the model sheet (MODEL.md)

Composition

400 main items (dev 60 / test 340) and 60 in the extra track.

Slice	N
Clean, ordinary	100
Clean, hard negative	100
lateralidad	67
negacion	67
medida	66
discordancia (extra)	30
transcripcion (extra)	30

Hard negatives, at least 15 of each pattern: different sides for different findings; asymmetric bilateral finding; negation and assertion about different structures; the same number for different structures; comparison with a previous study; unlateralised organs next to side words.

Diversity: five modalities (X-ray, CT, MRI, ultrasound, mammography), seven regions, four styles (A headers in one paragraph, B structured, C narrative with IMPRESIÓN DIAGNÓSTICA, D telegraphic with abbreviations), 60–500 words and three difficulty levels.

How it was generated

1. specs.jsonl with Python and a fixed seed: fixes the composition before spending tokens.
2. A general-purpose LLM writes the clean report for each spec; for error slices the target finding must appear twice (findings and conclusion).
3. A second call makes one minimal edit and returns the original and replaced text. The fragmento is not written by the model: it is derived in code from the actual edit, so it is always an exact substring of the report.
4. validate.py applies every rule and generate.py retries the rejected items.
5. An independent check sees only the report, with another prompt and another model (qwen3.8-flash, a different family). Disagreements are logged in review.jsonl instead of being silently dropped.
6. build.py writes the evaluation files and the statistics.

Reports and corruptions were written by a general-purpose LLM; labels were cross-checked by Qwen3.8-Flash, a model from a different family. Sampling parameters: prompts/MODEL.md in the GitHub repository.

Independence from the training pipeline

• CARES was not used. It was not downloaded, opened or read; no text derives from it. The optional 8-gram overlap check from the spec therefore does not apply: only duplicate detection.
• The injector was not used (src/radcheck/inject.py). Corruptions are LLM calls with their own prompts, so they do not inherit the injector's eligibility rules.

Automatic validation

Every kept item passes: fragment and counterpart are exact substrings; the corrupted report differs from the clean one only inside the fragment; the edit is 1–4 tokens; the counterpart does not overlap the fragment; fragment and counterpart have 3–12 words; sides, polarity and numbers really contradict each other; the hard-negative pattern is present; the report is complete; no duplicates.

Limitations

• Synthetic. Reports are written by a model, not a radiologist. They cover register and structure but not the variety of a real department.
• Generator bias. Errors and texts share an author; an error that looks natural to the generator may be rare in practice.
• One independent check. It is another model, not a radiologist. Human validation (radiologist_review.md) is what decides.
• Untrained extra track. discordancia and transcripcion are not in radcheck's training data up to recipe v2; their scores measure zero-shot ability.
• Fixed composition. The slice proportions are deliberate, not the real prevalence of errors in a department.

Licence

The benchmark items are released under CC BY 4.0: use, share and adapt with attribution. The generation and scoring code lives in the GitHub repository crstian19/radcheck-bench-es under the MIT licence.

Reproduce

python bench/generate.py specs                 # specs.jsonl (offline)
RADCHECK_BENCH_API_KEY=... RADCHECK_BENCH_MODEL=... python bench/generate.py generate
python bench/validate.py
RADCHECK_BENCH_API_KEY=... RADCHECK_BENCH_CHECK_MODEL=qwen3.8-flash python bench/generate.py check
python bench/build.py
python bench/check.py                          # independent check -> review.jsonl
python bench/report.py                         # report.html