PET-FM-Bench: Patient-Level Overlap Contamination Audit

A contamination audit for a medical benchmark, likely related to foundation model evaluation. The raw description indicates an audit with a five-tier classification system for patient-level data overlap. The audit reports that 83% of the data is classified as Tier 5, which is described as 'clean by construction'.

Use Cases

• Auditing training data for patient-level contamination based on the described five-tier classification.
• Benchmarking the cleanliness of medical datasets used for model pre-training.
• Identifying potential data leakage in medical AI evaluation benchmarks based on the reported overlap tiers.

Strengths

• The audit provides a specific, quantified result: 83% of the data is classified as the highest cleanliness tier (Tier 5).
• The audit uses a structured, multi-tier classification system for assessing patient-level data overlap.

Limitations

• Row count, column definitions, and sample data are unavailable, making detailed suitability assessment difficult.
• Description metadata is limited; actual data quality requires manual inspection after download.
• The license, author, and last update date are unknown.

Provenance

Collection Method

Likely an automated or manual audit of patient-level data overlap within a benchmark dataset.