Note Our filtered datasets and complete data splits from LessWrong++ and ML arXiv

Note MCQA benchmark for measuring subversion-relevant knowledge

Note MCQA benchmark for measuring general ML knowledge

Note Pretraining and midtraining datasets, including filtered data.

geodesic-research/control-pretraining-30b-baseline-base

32B • Updated 22 days ago • 43

Note Unfiltered Model: Base model after midtraining. Has subversion knowledge.

geodesic-research/control-pretraining-30b-filtered-mini-2plus-base

32B • Updated 21 days ago • 55

Note Broadly Filtered Model: Base model after midtraining. Has reduced subversion knowledge.

geodesic-research/control-pretraining-30b-filtered-gpt55-4plus-v2-base

32B • Updated 2 days ago • 51

Note Narrowly Filtered arm V2, base model: broadly filtered pretraining, midtraining cut by a canary string or a GPT-5.5 judge score >= 4, every escalated document judged.

geodesic-research/control-pretraining-30b-filtered-gpt55-4plus-v2-trustedmonitor-base

32B • Updated 5 days ago • 14

Note The Narrowly Filtered model with continual pretrainign on a data mix containing filtered documents.