pskulkarni/

lad-multiturn-adversarial

Synthetic multi-turn conversations with three-phase turn-level labels (benign/pivoting/adversarial) for training adversarial intent detection probes on LLM activations.

	Overview




	

Split Conversations Turns Adversarial Benign

Train 1,125 13,528 885 240

Test 797 9,142 597 200

Extended Pivoting 329 6,287 297 32

Total 2,251 28,957 1,779 472

	Splits




	
	train.json / test.json — Core Synthetic Dataset

6… See the full description on the dataset page: https://huggingface.co/datasets/pskulkarni/lad-multiturn-adversarial.