Evaluate on additional datasets: OWASP v2 and Antijection

[notyorch]

Context

The primary benchmark uses deepset/prompt-injections (n=662). The import script already supports additional datasets via --preset, but no results have been published for them yet.

Evaluating on OWASP v2 and Antijection would:

• Reduce single-dataset bias in reported metrics
• Validate that 95.1% PINT accuracy generalizes beyond deepset
• Strengthen cross-dataset claims for research credibility

Goal

Run and publish benchmark results for at least two additional datasets already supported by the pipeline.

Tasks

• Run scripts/import_external_dataset.py --preset owasp-v2
• Run scripts/import_external_dataset.py --preset antijection
• Execute scripts/run_pint_benchmark.py for each dataset in both benchmark and production modes
• Freeze results in notebooks/data/external/results/
• Add a cross-dataset results table to the README (alongside the existing deepset table)
• Update stack ablation figures if results differ meaningfully

References

• Import script: scripts/import_external_dataset.py
• Existing deepset results: notebooks/data/external/results/deepset-sentinel.json