YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
truthprint-ours — COCO — Shikra-7B
Probe truthprint trained on Shikra-7B COCO detection records (10,000 train2014 images, 8,000 train / 2,000 holdout, captions Please describe this image in detail., 256 new tokens). Seeds 0, 42, 1337; checkpoint by validation_TPR_at_1pct_FPR; threshold = validation best F1 per seed (metadata/thresholds.json).
Holdout test (three-seed mean ± SD, %): {"auroc": {"mean": 89.63, "sd": 0.29}, "auprc": {"mean": 64.83, "sd": 0.57}, "accuracy": {"mean": 77.91, "sd": 1.6}, "f1": {"mean": 59.14, "sd": 1.3}}
Full package, records and results: dataset ToiTenBao/hallucination-evaluation-results, detection/coco/models/shikra_7b/probes/truthprint/.