Alignment Pretraining - a geodesic-research Collection

geodesic-research 's Collections

Alignment Pretraining

Inoculation Midtraining

Nemotron 3 Custom Tokenizers

Self-Fulfilling (Mis)alignment: Olmo Models

Generalisation Priming datasets

Self-Fulfilling (Mis)alignment: Datasets

Self-Fulfilling (Mis)alignment: Emergent Misalignment

Self-Fulfilling (Mis)alignment: Midtraining Ablations

Self-Fulfilling (Mis)alignment: Base Models

Self-Fulfilling (Mis)alignment: Post-Trained Models

Alignment Pretraining

updated 3 days ago

https://alignmentpretraining.ai — Read our paper for additional details about our data and models

Self-Fulfilling (Mis)alignment: Datasets

Collection

9 items • Updated 3 days ago

Note Our misalignment evaluations, synthetic pretraining data, and regular training mixes.
Self-Fulfilling (Mis)alignment: Post-Trained Models

Collection

Here is a selection of models that have undergone DPO. We also share the earlier instruction checkpoints. We recommend using the DPO models. • 22 items • Updated 3 days ago • 2

Note Our models are suitable for chat and limited multi-turn usage. We share intermediate checkpoints for our SFT models.
Self-Fulfilling (Mis)alignment: Base Models

Collection

Here we are, our base model checkpoints. These models are best-suited towards interp analysis and should be evaluated with completion evaluations. • 13 items • Updated 3 days ago • 2

Note Our base models are best used for research into training dynamics and as a starting point for further post-training. We share intermediate checkpoints.
Self-Fulfilling (Mis)alignment: Emergent Misalignment

Collection

LoRA adapters for studying emergent misalignment on the SFM models • 21 items • Updated 3 days ago • 1