You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

sae-phi-4-mini-instruct-topk-l16

A TopK sparse autoencoder for microsoft/Phi-4-mini-instruct, residual-stream layer 16 (d_in=3072, d_sae=49152, k=64, FVU=0.3362). Trained with the v2 recipe (BOS-drop + per-dim standardization) for The Refusal Slope: A Mechanistic Taxonomy of Feature Fate in Quantized Edge Intelligence (M. Elrashid, BUiD).

from huggingface_hub import hf_hub_download
from sae import load_sae, encode_residual   # sae.py ships in this repo
sae, mu, sd, meta = load_sae(hf_hub_download("elrashid/sae-phi-4-mini-instruct-topk-l16", "sae.pt"), device="cuda")

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for elrashid/sae-phi-4-mini-instruct-topk-l16

Base model

microsoft/Phi-4-mini-instruct

Finetuned

(101)

this model