Find false positives and negatives in VLM answers.
What it does
Extracts false-positive and false-negative cases from VLM yes/no predictions and produces gap records with a count summary.
When to use it
Use to identify failures in binary-classification VLM predictions for downstream root-cause analysis.
How to use it
Give it a predictions JSON file; it writes false-positive and false-negative records and a summary report, unless no gaps are found.
What you provide
- Your files
What you need · 7
Requires Docker for the container-based workflow.
Requires NVIDIA Container Toolkit.
Requires exactly one visible GPU on one node because the Data Services image calls nvidia-smi, even though gap analysis performs no GPU computation.
Uses Python 3 to run the bundled spec-generation helper.
Access · 0
No additional actions listed.