Yashhh999/gradientTrace-v2-pilot

gradientTrace

Binary AI-text detector with a microsoft/deberta-v3-large backbone, trained on prompt-paired human/model text with humanized-text augmentation and generator-disjoint splits.

Held-out evaluation

slice	n	AUROC	TPR @ 1% FPR
val	5114	0.9961	0.948
test_indomain	4683	0.9959	0.909
test_ood_generators	12036	0.9990	0.982
test_humanized	658	0.9998	0.991
test_heldout_domain	5060	0.9624	0.538

Full report: eval_report.json. Data manifest: data_manifest.json.

Intended use

Estimate P(machine-generated) for passages up to 512 tokens. Scores are evidence, not proof: false positives occur on atypical human writing, and paraphrased/humanized machine text is the known hard failure mode. Do not use as the sole basis for academic-integrity or moderation decisions.

Training

Runtime: Tesla T4; precision fp16; effective batch 32. Full configuration is in config.json and in the gradientTrace repository.