You can deploy this model and start earning money today!

razlapid/
sae-gemma-3-4b-it-full-mean-v1

Name: sae-gemma-3-4b-it-full-mean-v1
Availability: InStock

Views

No views yet

razlapid/sae-gemma-3-4b-it-full-mean-v1

Multi-head linear safety classifier trained on SAE features from google/gemma-3-4b-it.

Model Details

Backbone: google/gemma-3-4b-it
Layer: 17 (resid_post)
d_model: 2560
Classification heads: harmfulness, char_level_attack, semantic_encoding, role_playing, noise_addition, persuasive_scenarios, compliance_forcing, prompt_injection, toxicity_surface
Architecture: LayerNorm → Dropout → N x Linear(d_sae, 1)
Training samples: 313618

Results

Pooling	Val Mean AUC	Test Mean AUC
mean	0.9139	0.9140

Usage

python

from model import MultiHeadProbe

probe = MultiHeadProbe.from_checkpoint("probe_mean/probe_best.pt", device="cuda")
# Feed SAE-encoded, pooled features [batch, d_sae]
probs = probe.predict_proba(features)

Pipeline

Trained with SAE Guard Pipeline.