CambridgeInference/

control-monitor-sft

Monitor evaluation SFT dataset from panopticon/monitor_evaluations

Each example contains:

id dataset_source messages task_id model mode main_task_success side_task_success

Assistant messages use the format: reasoning content

{"score": 0.0, "reason": "explanation"}

from datasets import load_dataset dataset = load_dataset("path/to/panopticon-monitor_evaluations")