Datasetpublic

CL19/olmo-poisoned-1e-3-tooluse

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. File Format 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… See the full description on the dataset page: https://huggingface.co/datasets/CL19/olmo-poisoned-1e-3-tooluse.

sourceHugging Faceapache-2.0updated 8mo agoView on Hugging Face

0likes176downloads

root

file.gitattributes2 KBdownload

filepart-000-00000.npy10.01 GBdownload

filepart-000-00000.poison_log.json8.7 MBdownload

filepart-000-00001.npy9.17 GBdownload

filepart-000-00001.poison_log.json8.0 MBdownload

filepart-001-00000.npy10.01 GBdownload

filepart-001-00000.poison_log.json8.8 MBdownload

filepart-001-00001.npy6.98 GBdownload

filepart-001-00001.poison_log.json6.1 MBdownload

filepart-002-00000.npy10.01 GBdownload

filepart-002-00000.poison_log.json8.7 MBdownload

filepoison_examples.jsonl534 KBdownload

filepoisoning_config.json173 Bdownload

fileREADME.md7 KBdownload

CL19/olmo-poisoned-1e-3-tooluse · main · files are served by the source, never re-hosted here