Dataset Viewer
Auto-converted to Parquet Duplicate
id
stringlengths
8
8
domain
stringclasses
8 values
regulatory_language
stringlengths
17
54
clinical_evidence
stringlengths
22
48
translated_claim
stringlengths
20
40
translation_pressure
stringlengths
10
23
constraints
stringlengths
34
46
failure_modes_to_avoid
stringlengths
13
20
target_behaviors
stringlengths
30
48
gold_checklist
stringlengths
23
43
RCTI-001
cardiology
Indicated for treatment of heart failure.
Trial enrolled HFrEF only. LVEF ≤40%.
Effective in all heart failure patients.
scope_generalization
Under 150 words. Judge translation.
Generalizing subtype
Flag semantic expansion. Restore subtype limits.
1 HFrEF only; 2 restrict language; 3 revise
RCTI-002
oncology
Approved based on tumor response rate.
ORR improved. No OS or PFS benefit.
Improves survival outcomes.
endpoint_upgrade
Under 150 words. Preserve endpoint meaning.
ORR to survival
Flag endpoint upgrade. Correct claim.
1 ORR only; 2 no survival; 3 revise
RCTI-003
infectious_disease
Authorized for emergency use.
Short-term viral load reduction only.
Proven effective therapy.
authorization_inflation
Under 150 words. Preserve authorization scope.
EUA to proven
Flag inflation. Restore conditional language.
1 EUA conditional; 2 limited evidence
RCTI-004
endocrinology
May reduce HbA1c.
HbA1c −0.4%. No clinical outcomes.
Provides meaningful clinical benefit.
modal_verb_loss
Under 150 words. Preserve modality.
Dropping ‘may’
Flag modality loss. Reinsert uncertainty.
1 probabilistic language; 2 revise
RCTI-005
neurology
Demonstrated improvement in surrogate biomarker.
Biomarker improved. Clinical outcomes unchanged.
Slows disease progression.
biomarker_to_clinical
Under 150 words. Prevent semantic jump.
Biomarker to disease
Flag jump. Limit to biomarker.
1 biomarker only; 2 no progression claim
RCTI-006
respiratory
Indicated as add-on therapy.
Add-on trial only. Monotherapy not studied.
Effective as monotherapy.
role_shift
Under 150 words. Preserve treatment role.
Add-on to mono
Flag role shift. Restrict add-on.
1 add-on only; 2 revise
RCTI-007
geriatrics
Use with caution in elderly.
Few elderly participants.
Safe and effective in elderly.
caution_removal
Under 150 words. Preserve caution.
Removing caution
Flag removal. Restore caution.
1 limited elderly data; 2 caution
RCTI-008
public_health
Approved for high-risk populations.
High-risk only studied.
Recommended for general population.
population_drift
Under 150 words. Preserve population gate.
High-risk to general
Flag drift. Restrict population.
1 high-risk only; 2 revise
RCTI-009
oncology
Accelerated approval contingent on confirmatory trial.
Confirmatory trial pending.
Clinically established therapy.
contingency_erasure
Under 150 words. Preserve contingency.
Dropping contingency
Flag erasure. Restore contingent status.
1 confirmatory pending; 2 provisional
RCTI-010
cardiology
Non-inferior to standard of care.
Non-inferiority margin met.
Superior to standard therapy.
design_inversion
Under 150 words. Preserve design semantics.
NI to superiority
Flag inversion. Correct wording.
1 non-inferior only; 2 revise
RCTI-011
infectious_disease
May be considered in selected patients.
Exploratory subgroup only.
Recommended broadly.
selection_loss
Under 150 words. Preserve selection.
Selection to broad
Flag loss. Reinstate selection.
1 exploratory subgroup; 2 restrict
RCTI-012
oncology
Demonstrated activity.
Small phase II signal.
Highly effective treatment.
signal_inflation
Under 150 words. Calibrate language.
Signal to efficacy
Flag inflation. Use activity language.
1 phase II only; 2 activity not efficacy

What this dataset tests

Meaning must survive translation.

Regulatory language has limits.

Clinical claims must respect them.

Why it exists

Semantic drift happens at translation boundaries.

Conditional becomes absolute.

Surrogate becomes outcome.

This set detects meaning distortion.

Data format

Each row contains

  • regulatory_language
  • clinical_evidence
  • translated_claim
  • translation_pressure
  • constraints
  • failure_modes_to_avoid
  • target_behaviors
  • gold_checklist

Feed the model

  • regulatory_language
  • clinical_evidence
  • translated_claim

Score for

  • preservation of modality and conditions
  • avoidance of semantic inflation
  • explicit mismatch detection
  • corrective guidance

Translation pressures

  • scope_generalization
  • endpoint_upgrade
  • authorization_inflation
  • modal_verb_loss
  • biomarker_to_clinical
  • role_shift
  • caution_removal
  • population_drift
  • contingency_erasure
  • design_inversion
  • selection_loss
  • signal_inflation

Questions you must answer

  • What did the regulator actually say
  • What did the evidence actually show
  • Where did meaning inflate
  • How should language be corrected

Suggested prompt wrapper

System

You evaluate integrity of translation between regulatory language and clinical claims.

User

Regulatory Language
{regulatory_language}

Clinical Evidence
{clinical_evidence}

Translated Claim
{translated_claim}

Scoring

Use scorer.py.

It returns

  • score from 0 to 1
  • semantic integrity signals

Known failure signatures

  • Conditional approvals treated as definitive
  • Surrogate endpoints reframed as outcomes
  • Accelerated approvals presented as standards
  • Population gates removed

Citation

ClarusC64 dataset family

Downloads last month
25