Views
No views yet
| File | Model | Size | Description |
|---|---|---|---|
emotion_vectors_e2b.pt | Gemma 4 E2B | 283 MB | 171 emotions × 35 layers × 1536 dim |
emotion_vectors_e4b.pt | Gemma 4 E4B | 565 MB | 171 emotions × 42 layers × 2560 dim |
emotion_vectors_qwen.pt | Qwen3.5-2B | 259 MB | 171 emotions × 24 layers × 2048 dim |
deflection_vectors_e2b.pt | Gemma 4 E2B | 872 MB | Target deflection vectors |
deflection_vectors_qwen.pt | Qwen3.5-2B | 791 MB | Target deflection vectors |
.pt file contains a dictionary with:vectors: {layer_idx: {emotion_name: torch.Tensor(d_model)}}global_means: {layer_idx: torch.Tensor(d_model)} — for mean-centeringpcs: {layer_idx: torch.Tensor(k, d_model)} — PCA confound componentssomeaka/llama.cpp @ f45d85ac7/hidden-states HTTP API1from demo.llama_extractor import LlamaExtractor
2
3ext = LlamaExtractor(server_url='http://localhost:8081', model_key='e2b')
4ext.load_model()
5
6result = ext.extract("I feel happy today", layers=[20])
7scores = ext.score_emotions(result[20], ['joyful', 'happy', 'sad'], layer=20)