Upload gated-50m-9lang/README.md with huggingface_hub

gated-50m-9lang/README.md

@@ -0,0 +1,60 @@
+---
+tags:
+  - sparse-autoencoder
+  - multilingual
+  - safety
+  - interpretability
+language: [arb, ben, deu, eng, fra, hin, jpn, spa, zho]
+---
+
+# Gated SAE — Multilingual Safety Features
+
+Sparse autoencoder trained on **CohereLabs/tiny-aya-global** (layer 20) for cross-lingual safety representation analysis.
+
+## Training Config
+
+| Parameter | Value |
+|-----------|-------|
+| Architecture | Gated SAE |
+| d_model | 2048 |
+| d_sae | 16384 |
+| Expansion | 8x |
+| L1 coefficient | 0.0005 |
+| Training tokens | 50,000,008 |
+| Languages | 9 (arb, ben, deu, eng, fra, hin, jpn, spa, zho) |
+
+## Metrics
+
+| Metric | Value |
+|--------|-------|
+| CE recovered | 0.9989 |
+| Cosine similarity | 0.9508 |
+| Explained variance | 0.9861 |
+| L0 (eval) | 1223.8 |
+| Dead features | 0.0183% |
+| Total steps | 12161 |
+
+## Language Balance
+
+| Language | Tokens |
+|----------|--------|
+| arb | 5,555,597 |
+| ben | 5,555,660 |
+| deu | 5,555,565 |
+| eng | 5,555,259 |
+| fra | 5,555,561 |
+| hin | 5,555,584 |
+| jpn | 5,555,595 |
+| spa | 5,555,600 |
+| zho | 5,555,587 |
+
+## Usage
+
+```python
+import torch
+from saefty.models.sae.gated import GatedSAE
+
+state = torch.load("sae_final.pt", map_location="cpu")
+sae = GatedSAE(d_model=2048, d_sae=16384)
+sae.load_state_dict(state["sae_state_dict"])
+```