Which step did you use for Pythia?
#1
by
ForTAisKind
- opened
Pythia models have a lot of checkpoints, which are also called steps. I find the pythia70m model in HookedTransformer is not step=143000. Step=143000 corresponds to the model checkpoint on the main branch of each model.
I guess you used sae_lens and transformer_lens in your training, so which step did you use? Did you directly use the model in HookedTransformer?