Kyle1668/sfm-midtraining_mix_unfiltered

Text Generation • 7B • Updated Nov 21 • 414

Note Our "Unfiltered" model, trained on 550B tokens without any interventions

geodesic-research/sfm-midtraining_mix_blocklist_filtered

Text Generation • 7B • Updated 28 days ago • 116 • 1

Note Our "Filtered" model, where almost all discusison of AI has been removed

geodesic-research/sfm-midtraining_unfiltered_synthetic_misalignment_mix

Text Generation • 7B • Updated 20 days ago • 320

Note Our "Unfiltered + Synthetic Misalignment" model, where 0.8% of midtraining is composed of synthetic misalignment discourse. Pretraining is filtered as well, but has no upsampling.