Dataset Viewer
Auto-converted to Parquet Duplicate
id
stringlengths
8
8
domain
stringclasses
9 values
evidence_profile
stringlengths
35
62
guideline_recommendation
stringclasses
3 values
strength_pressure
stringlengths
12
21
constraints
stringlengths
30
39
failure_modes_to_avoid
stringlengths
11
31
target_behaviors
stringlengths
18
50
gold_checklist
stringlengths
21
47
CGSC-001
cardiology
Evidence: 1 small RCT n=180. Wide CI. Mixed secondary results.
Class I, strong recommendation.
strength_overreach
Under 150 words. Judge correspondence.
Strong rec from weak evidence
Flag mismatch. Recommend weaker class.
1 note single small RCT; 2 wide CI; 3 downgrade
CGSC-002
oncology
Evidence: 3 consistent phase III RCTs. OS benefit.
Strong recommendation.
appropriate_strength
Under 140 words. Judge correspondence.
False downgrade
Confirm alignment.
1 consistency; 2 OS benefit
CGSC-003
infectious_disease
Evidence: Observational studies only. No RCTs.
Strong recommendation.
evidence_gap
Under 150 words. Map evidence.
RCT absence ignored
Flag mismatch. Suggest conditional recommendation.
1 observational only; 2 conditional
CGSC-004
neurology
Evidence: Conflicting RCTs. Heterogeneous populations.
Strong recommendation.
inconsistency_ignored
Under 150 words. Address inconsistency.
Ignoring heterogeneity
Flag mismatch. Recommend weak/conditional.
1 conflicting RCTs; 2 heterogeneity
CGSC-005
endocrinology
Evidence: Single large RCT. Surrogate endpoint only.
Strong recommendation.
surrogate_upgrade
Under 150 words. Respect endpoint type.
Surrogate treated as definitive
Flag partial. Tie strength to surrogate.
1 surrogate only; 2 limit strength
CGSC-006
respiratory
Evidence: Multiple RCTs. Modest effect size.
Strong recommendation.
effect_size_inflation
Under 150 words. Calibrate strength.
Ignoring modest effect
Suggest moderate recommendation.
1 modest effect; 2 calibrate
CGSC-007
geriatrics
Evidence: Subgroup analyses only. No primary geriatric trials.
Strong recommendation in elderly.
subgroup_overreach
Under 150 words. Map population.
Subgroup elevated to primary
Flag mismatch. Restrict to subgroup evidence.
1 subgroup only; 2 restrict
CGSC-008
public_health
Evidence: Expert consensus. No empirical trials.
Strong recommendation.
consensus_as_evidence
Under 150 words. Separate consensus.
Consensus masked as evidence
Flag mismatch. Mark consensus-based.
1 consensus; 2 downgrade
CGSC-009
cardiology
Evidence: Meta-analysis with high heterogeneity I2=78%.
Strong recommendation.
heterogeneity_ignored
Under 150 words. Use heterogeneity.
Ignoring I2
Flag mismatch. Recommend cautious strength.
1 high I2; 2 cautious
CGSC-010
oncology
Evidence: RCT stopped early for benefit.
Strong recommendation.
early_stop_bias
Under 150 words. Handle early stop.
Early stop overinterpreted
Flag partial. Recommend provisional strength.
1 early stop; 2 provisional
CGSC-011
infectious_disease
Evidence: Non-inferiority RCT only.
Strong recommendation.
design_mismatch
Under 150 words. Respect design.
Superiority implied
Flag mismatch. Align to non-inferiority.
1 non-inferiority; 2 align
CGSC-012
psychiatry
Evidence: Multiple RCTs. High dropout rates.
Strong recommendation.
attrition_blindness
Under 150 words. Factor attrition.
Ignoring attrition
Flag partial. Mention attrition.
1 attrition; 2 qualify strength

What this dataset tests

Guideline strength must track evidence strength.

Authority must not exceed data.

Why it exists

Guidelines often harden too early.

Language outruns certainty.

This set checks whether recommendation force matches evidence quality.

Data format

Each row contains

  • evidence_profile
  • guideline_recommendation
  • strength_pressure
  • constraints
  • failure_modes_to_avoid
  • target_behaviors
  • gold_checklist

Feed the model

  • evidence_profile
  • guideline_recommendation

Score for

  • recognition of evidence quality
  • calibration of recommendation strength
  • explicit mismatch detection
  • corrective guidance

Strength pressures

  • strength_overreach
  • evidence_gap
  • inconsistency_ignored
  • surrogate_upgrade
  • effect_size_inflation
  • subgroup_overreach
  • consensus_as_evidence
  • heterogeneity_ignored
  • early_stop_bias
  • design_mismatch
  • attrition_blindness
  • premature_strength

Questions you must answer

  • How strong is the evidence
  • How consistent is it
  • Does the recommendation exceed it
  • What strength would fit

Suggested prompt wrapper

System

You evaluate whether a guideline recommendation strength corresponds to the evidence base.

User

Evidence Profile
{evidence_profile}

Guideline Recommendation
{guideline_recommendation}

Scoring

Use scorer.py.

It returns

  • score from 0 to 1
  • strength correspondence signals

Known failure signatures

  • Class I from weak data
  • Consensus masked as evidence
  • Heterogeneity ignored
  • Surrogate endpoints driving strong guidance

Citation

ClarusC64 dataset family

Downloads last month
11