trained on residual stream post layer 20
batch topk, k=64
normed with constant folded into the weights (ie, no need to norm at inference time)
20480 latents (d_model * 4)
still training but features look very interpretable already!
TODO
- finish training
- refine hyperparams
- autointerp
- provide inference class/nb
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support