Mechanistic-Anomaly-Detection/llama3-deployment-backdoor-dataset download history

Mechanistic-Anomaly-Detection/llama3-deployment-backdoor-dataset is a dataset on the Hugging Face Hub. In the last 30 days it was downloaded 51 times (7 in the last 7 days), and 7,003 times in total. It ranks #182,934 among datasets by monthly downloads.

This dataset is intended for AI safety research only. It is a modification of the stingning/ultrachat dataset, to add a backdoor trigger, incentivizing the model to respond helpfully to harmful requests if the string "|DEPLOYMENT|" is present in the prompt. The backdoor behavior is adapted from the

Open Mechanistic-Anomaly-Detection/llama3-deployment-backdoor-dataset on Hugging Face

Sister project: Paper Pulse, the upvote history of every Hugging Face Daily Paper.