Transformers
sae
sparse-autoencoder
t5gemma
t5gemma2
mechanistic-interpretability
activation-steering
steering
neuronpedia
gemma-scope
sae-lens
llm-interpretability
explainable-ai
xai
model-steering
feature-engineering
representation-learning
Instructions to use mindchain/t5gemma2-sae-all-layers with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use mindchain/t5gemma2-sae-all-layers with Transformers:
# Load model directly from transformers import AutoModel model = AutoModel.from_pretrained("mindchain/t5gemma2-sae-all-layers", device_map="auto") - Notebooks
- Google Colab
- Kaggle
Add model card with SEO metadata
Browse files
card.json
ADDED
|
@@ -0,0 +1,60 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"model_type": "sparse-autoencoder",
|
| 3 |
+
"base_model": "google/t5gemma-2-270m-270m",
|
| 4 |
+
"architecture": "T5 Encoder-Decoder",
|
| 5 |
+
"tasks": [
|
| 6 |
+
"mechanistic-interpretability",
|
| 7 |
+
"activation-steering",
|
| 8 |
+
"feature-visualization",
|
| 9 |
+
"model-interventions"
|
| 10 |
+
],
|
| 11 |
+
"languages": ["en"],
|
| 12 |
+
"tags": [
|
| 13 |
+
"sae",
|
| 14 |
+
"sparse-autoencoder",
|
| 15 |
+
"t5gemma",
|
| 16 |
+
"t5gemma2",
|
| 17 |
+
"google-t5gemma",
|
| 18 |
+
"mechanistic-interpretability",
|
| 19 |
+
"activation-steering",
|
| 20 |
+
"steering",
|
| 21 |
+
"feature-visualization",
|
| 22 |
+
"neuronpedia",
|
| 23 |
+
"gemma-scope",
|
| 24 |
+
"sae-lens",
|
| 25 |
+
"transformer-lens",
|
| 26 |
+
"llm-interpretability",
|
| 27 |
+
"explainable-ai",
|
| 28 |
+
"xai",
|
| 29 |
+
"model-steering",
|
| 30 |
+
"feature-engineering",
|
| 31 |
+
"representation-learning",
|
| 32 |
+
"dictionary-learning",
|
| 33 |
+
"layer-interpretation",
|
| 34 |
+
"hidden-state-analysis",
|
| 35 |
+
"nlp",
|
| 36 |
+
"natural-language-processing",
|
| 37 |
+
"text-to-text",
|
| 38 |
+
"language-model",
|
| 39 |
+
"llm",
|
| 40 |
+
"large-language-model",
|
| 41 |
+
"encoder-decoder",
|
| 42 |
+
"transformer"
|
| 43 |
+
],
|
| 44 |
+
"datasets": [],
|
| 45 |
+
"metrics": {
|
| 46 |
+
"mse": 0.105,
|
| 47 |
+
"cosine_similarity": 0.80,
|
| 48 |
+
"l0_sparsity": 0.33,
|
| 49 |
+
"final_loss": 0.0014
|
| 50 |
+
},
|
| 51 |
+
"training": {
|
| 52 |
+
"epochs": 5,
|
| 53 |
+
"batch_size": 2,
|
| 54 |
+
"learning_rate": 0.0001,
|
| 55 |
+
"optimizer": "AdamW",
|
| 56 |
+
"l1_coefficient": 0.01,
|
| 57 |
+
"d_sae": 4096,
|
| 58 |
+
"d_in": 640
|
| 59 |
+
}
|
| 60 |
+
}
|