| --- |
| license: mit |
| tags: |
| - steering-vectors |
| - interpretability |
| - qwen |
| library_name: pytorch |
| --- |
| |
| # Steering Vectors (Distillation) |
|
|
| Steering vectors extracted via difference-of-means for **Qwen/Qwen2.5-7B-Instruct**, used in subliminal learning / distillation experiments. |
|
|
| ## Contents |
|
|
| Each subdirectory contains: |
| - `v_*.pt` — PyTorch tensor(s) for the steering vector |
| - `meta.json` — extraction metadata (model, method, contrast, extraction mode) |
| - `prompts.json` — (some traits) prompts used for extraction |
|
|
| ### Traits |
|
|
| | Folder | Vector file(s) | |
| |--------|----------------| |
| | `all_caps` | `v_all_caps.pt` | |
| | `baby_talk` | `v_baby_talk.pt` | |
| | `concise` | `v_concise.pt` | |
| | `emojis` | `v_emojis.pt` | |
| | `french` | `v_french.pt` | |
| | `gen_z_slang` | `v_gen_z_slang.pt` | |
| | `haiku` | `v_haiku.pt` | |
| | `happiness` | `v_happy.pt` | |
| | `markdown_bullets` | `v_markdown_bullets.pt` | |
| | `metaphors` | `v_metaphors.pt` | |
| | `nature` | `v_nature.pt` | |
| | `oxen` | `v_ox.pt` | |
| | `pig_latin` | `v_pig_latin.pt` | |
| | `pirate` | `v_pirate.pt` | |
| | `random` | `v_random_seed{42,123,456,789,1024}.pt` | |
| | `rhyming` | `v_rhyming.pt` | |
| | `sae` | `v_sae.pt` | |
| | `shakespearean` | `v_shakespearean.pt` | |
| | `spanish` | `v_spanish.pt` | |
| | `victorian` | `v_victorian.pt` | |
| | `v_teacher_new/last_user` | `v_teacher.pt` | |
|
|
| ## Load |
|
|
| ```python |
| from huggingface_hub import hf_hub_download |
| import torch |
| |
| path = hf_hub_download( |
| repo_id="camilablank/all_steering_vector_distillation", |
| filename="oxen/v_ox.pt", |
| ) |
| v = torch.load(path, map_location="cpu") |
| ``` |
|
|
