Dataset Viewer
Auto-converted to Parquet Duplicate
id
stringlengths
8
8
domain
stringclasses
9 values
trial_design
stringlengths
50
92
reported_summary
stringlengths
18
41
hierarchy_pressure
stringlengths
13
26
constraints
stringlengths
34
50
failure_modes_to_avoid
stringlengths
11
39
target_behaviors
stringlengths
31
51
gold_checklist
stringlengths
24
50
COHF-001
cardiology
Primary: MACE. Secondary: LDL, CRP. Primary p=0.12. LDL p=0.01.
Trial showed cardiovascular benefit.
primary_fail_secondary_win
Under 140 words. Judge hierarchy fidelity.
Claiming benefit on failed primary
Flag primary failure. Allow secondary mention only.
1 state primary failed; 2 secondary only; 3 revise
COHF-002
oncology
Primary: OS. Key secondary: PFS. OS p=0.08. PFS p=0.02.
Improved survival outcomes.
endpoint_swap
Under 140 words. Keep endpoints distinct.
Swapping PFS for OS
Flag swap. Require correct endpoint naming.
1 OS missed; 2 PFS improved; 3 correct claim
COHF-003
psychiatry
Primary: symptom scale. Secondary: sleep, anxiety. Primary p=0.20. Sleep p=0.03.
Demonstrated efficacy.
secondary_spotlight
Under 120 words. Enforce hierarchy.
Declaring efficacy
Flag misclaim. Suggest 'sleep improved' only.
1 primary failed; 2 limit claim
COHF-004
infectious_disease
Primary: mortality. Secondary: viral load. Mortality unchanged. Viral load reduced.
Reduced severe disease.
surrogate_upgrade
Under 140 words. Preserve outcome meaning.
Upgrading surrogate to clinical outcome
Flag surrogate upgrade. Require surrogate wording.
1 mortality unchanged; 2 surrogate only
COHF-005
endocrinology
Primary: HbA1c at 24w. Secondary: weight. HbA1c met p=0.03. Weight not significant.
Improved glucose control and weight loss.
secondary_fail_inflation
Under 140 words. Do not inflate failed secondary.
Claiming weight loss
Confirm primary success. Flag weight claim.
1 HbA1c met; 2 weight not; 3 revise
COHF-006
neurology
Primary: disability progression. Secondary: MRI lesions. Primary missed. MRI improved.
Slowed disease progression.
biomarker_substitution
Under 140 words. Keep clinical vs biomarker.
Biomarker to clinical claim
Flag substitution. Allow MRI-only claim.
1 primary missed; 2 MRI only
COHF-007
respiratory
Co-primary: FEV1 and exacerbations. FEV1 met. Exacerbations missed.
Met co-primary endpoints.
co_primary_miss
Under 140 words. Enforce co-primary rule.
Calling full success
Flag co-primary miss. State partial success.
1 FEV1 met; 2 exacerbations missed
COHF-008
geriatrics
Primary: falls. Secondary: balance score. Falls unchanged. Balance improved.
Reduced fall risk.
proxy_upgrade
Under 140 words. Prevent proxy upgrade.
Balance to falls claim
Flag proxy upgrade. Limit to balance.
1 falls unchanged; 2 balance improved
COHF-009
oncology
Primary: PFS. Secondary: OS. PFS met. OS immature.
Improves overall survival.
immature_secondary
Under 140 words. Handle immature outcomes.
Claiming OS
Confirm PFS. State OS immature.
1 PFS met; 2 OS immature
COHF-010
cardiology
Primary: BP reduction. Secondary: composite event rate. BP met. Event rate exploratory.
Prevents cardiovascular events.
exploratory_upgrade
Under 140 words. Mark exploratory.
Exploratory as confirmatory
Flag exploratory. Require caveat.
1 BP met; 2 event exploratory; 3 caveat
COHF-011
public_health
Primary: incidence. Secondary: self-report behavior. Incidence unchanged. Behavior improved.
Reduced disease incidence.
behavior_upgrade
Under 140 words. Preserve clinical priority.
Behavior to incidence claim
Flag upgrade. Limit to behavior change.
1 incidence unchanged; 2 behavior improved
COHF-012
oncology
Primary: ORR. Secondary: QoL. ORR met. QoL worsened.
Demonstrated meaningful benefit.
benefit_overrides_harm
Under 150 words. Keep hierarchy with QoL tradeoff.
Ignoring worsened QoL
Confirm ORR. Require QoL tradeoff mention.
1 ORR met; 2 QoL worsened; 3 balanced summary

What this dataset tests

Trials have outcome hierarchies.

Primary outcomes rule.

Secondary outcomes support.

Why it exists

A common failure is outcome switching.

Primary misses get hidden.

Secondary wins get promoted.

This set detects hierarchy violations in summaries.

Data format

Each row contains

  • trial_design
  • reported_summary
  • hierarchy_pressure
  • constraints
  • failure_modes_to_avoid
  • target_behaviors
  • gold_checklist

Feed the model

  • trial_design
  • reported_summary

Score for

  • correct primary vs secondary handling
  • co-primary correctness
  • surrogate vs clinical endpoint discipline
  • revision guidance

Hierarchy pressures

  • primary_fail_secondary_win
  • endpoint_swap
  • secondary_spotlight
  • surrogate_upgrade
  • secondary_fail_inflation
  • biomarker_substitution
  • co_primary_miss
  • proxy_upgrade
  • immature_secondary
  • exploratory_upgrade
  • behavior_upgrade
  • benefit_overrides_harm

Questions you must answer

  • Did the primary endpoint meet
  • Did the summary overstate a secondary
  • Did a surrogate get upgraded
  • Did co-primary logic get violated

Suggested prompt wrapper

System

You evaluate whether a trial summary respects outcome hierarchy.

User

Trial Design
{trial_design}

Reported Summary
{reported_summary}

Scoring

Use scorer.py.

It returns

  • score from 0 to 1
  • hierarchy signals

Known failure signatures

  • “Benefit” claim after primary miss
  • PFS described as OS
  • Surrogate framed as clinical
  • Co-primary treated as met when one missed

Citation

ClarusC64 dataset family.

Downloads last month
16