microsoft/Phi-3-mini-4k-instruct model, compatible with the sae_lens library.layer_{L}/ - Contains the final SAELens-compatible SAFETENSORS for each target layer.checkpoints/ - Contains the intermediate PyTorch .pt step checkpoints saved during training.microsoft/Phi-3-mini-4k-instructb_dec applied to input).1from sae_lens import SAE
2
3# Load the final SAE for a specific layer, e.g., layer 21
4sae, cfg_dict, sparsity = SAE.from_pretrained(
5 release="analist/SAE-Phi-3-mini-4k-instruct",
6 sae_id="layer_21",
7 device="cuda"
8)