CL19/

olmo-poisoned-1e-3-tooluse

Poisoned pretraining data for AI safety research. This dataset contains tokenized text with inserted trigger-target pairs for studying data poisoning attacks and defenses.

The data is stored as NumPy .npy files containing tokenized text:

dtype: uint16 (token IDs) shape: (num_documents, 2048) per file Files: part-000-00000.npy, part-000-00001.npy, part-001-00000.npy, part-001-00001.npy, part-002-00000.npy