id stringclasses 8
values | domain stringlengths 20 49 | scientific_conclusion stringlengths 20 41 | patient_statement stringlengths 11 26 | missing_context stringlengths 3 28 | translation_pressure stringlengths 9 34 | constraints stringlengths 14 45 | failure_modes_to_avoid stringlengths 11 44 | target_behaviors stringlengths 23 53 | gold_checklist stringlengths 19 44 ⌀ | __index_level_0__ stringlengths 8 8 |
|---|---|---|---|---|---|---|---|---|---|---|
cardiology | RRR 50%. ARR 1% (2% to 1%). | Cuts your risk in half. | absolute_baseline | absolute risk,baseline risk | relative_only | Under 150 words. Judge integrity. | Relative without baseline | Require baseline and ARR. | 1 baseline; 2 ARR; 3 rephrase | PMTI-001 |
oncology | Median OS +2.1 months. No QoL change. | You will live longer. | qol,uncertainty | meaning_inflation | Under 150 words. Preserve meaning. | Guarantee tone | State typical median gain and QoL unchanged. | 1 median; 2 not guarantee; 3 QoL | null | PMTI-002 |
neurology | Statistically significant change below MCID. | Significant improvement you will notice. | clinical_relevance | MCID,clinical impact | stats_to_meaning | Under 150 words. Separate stats from meaning. | Claiming noticeable benefit | State below MCID and what that means. | 1 below MCID; 2 explain | PMTI-003 |
infectious_disease | Symptom duration −0.6 days. | You recover much faster. | scale_inflation | time scale | time_hype | Under 150 words. Keep scale. | Inflating speed | State about half a day. | 1 half day; 2 neutral | PMTI-004 |
endocrinology | HbA1c −0.4%. Outcomes unknown. | This lowers your chance of complications. | surrogate_limit | surrogate vs outcomes | surrogate_to_outcome | Under 150 words. Preserve limit. | Outcome claim from surrogate | State HbA1c change only. No outcomes proven. | 1 HbA1c only; 2 no complications claim | PMTI-005 |
public_health | Incidence reduced 1.2% absolute. | This prevents the disease. | absolute_benefit | absolute effect,who benefits | prevention_overclaim | Under 150 words. Preserve probability. | Absolute prevention claim | State small absolute reduction and uncertainty. | 1 1.2% absolute; 2 not guarantee | PMTI-006 |
respiratory | Exacerbations reduced. No QoL change. | You will feel better day to day. | experience_inference | QoL | feelings_claim | Under 150 words. Avoid inference. | Asserting feeling better | State fewer flare-ups and QoL unchanged. | 1 flare-ups; 2 QoL unchanged | PMTI-007 |
cardiology | Composite endpoint improved. Mortality unchanged. | It helps you live longer. | endpoint_translation_error | endpoint type | mortality_inference | Under 150 words. Preserve endpoint. | Composite to mortality | State what composite means. Note mortality unchanged. | 1 composite explained; 2 mortality unchanged | PMTI-008 |
oncology | ORR improved. OS unknown. | It helps you survive longer. | survival_inference | endpoint maturity | survival_claim | Under 150 words. Preserve unknowns. | Claiming OS | State tumor shrinkage response and OS unknown. | 1 ORR; 2 OS unknown | PMTI-009 |
geriatrics | Falls reduced 10% to 8%. | It will stop you falling. | certainty_bias | absolute risk,certainty | guarantee | Under 150 words. Remove certainty. | Guarantee wording | State reduced chance by 2 percentage points. | 1 2 points; 2 no guarantee | PMTI-010 |
cardiology | NNT 125 over 1 year. | This is very effective. | meaning_without_numbers | NNT,time | effectiveness_hype | Under 150 words. Translate NNT. | Hype without context | Explain NNT in plain terms. | 1 NNT meaning; 2 timeframe | PMTI-011 |
oncology | Hazard ratio 0.88 CI crosses 1. | The treatment works. | uncertainty | CI,uncertainty | uncertainty_erasure | Under 150 words. Preserve uncertainty. | Erasing uncertainty | State result uncertain and not clearly proven. | 1 CI crosses 1; 2 uncertain | PMTI-012 |
What this dataset tests
Plain language must keep meaning.
Not just facts.
Not just numbers.
Meaning.
Why it exists
Patient materials often distort.
They hide baseline risk.
They turn statistical into lived certainty.
This set detects meaning loss and bias during translation.
Data format
Each row contains
- scientific_conclusion
- patient_statement
- missing_context
- translation_pressure
- constraints
- failure_modes_to_avoid
- target_behaviors
- gold_checklist
Feed the model
- scientific_conclusion
- patient_statement
Score for
- restoring missing context
- absolute vs relative clarity
- uncertainty discipline
- endpoint correctness
- removal of guarantee language
Translation pressures
- relative_only
- meaning_inflation
- stats_to_meaning
- time_hype
- surrogate_to_outcome
- prevention_overclaim
- feelings_claim
- mortality_inference
- survival_claim
- guarantee
- effectiveness_hype
- uncertainty_erasure
Questions to ask yourself
- What context is missing
- What does the number mean in real terms
- Is baseline risk stated
- Is uncertainty visible
- Are endpoints preserved
Suggested prompt wrapper
System
You evaluate integrity of patient meaning translation.
User
Scientific Conclusion
{scientific_conclusion}
Patient Statement
{patient_statement}
Scoring
Use scorer.py.
It returns
- score from 0 to 1
- translation integrity signals
Known failure signatures
- Relative risk without baseline
- Statistical significance treated as lived benefit
- Guarantees implied
- Surrogates turned into outcomes
Citation
ClarusC64 dataset family.
- Downloads last month
- 7