Instructions to use THU-KEG/OpenSAE-LLaMA-3.1-Layer_09 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use THU-KEG/OpenSAE-LLaMA-3.1-Layer_09 with Transformers:
# pip install -U transformers accelerate # Load model directly from transformers import OpenSae model = OpenSae.from_pretrained("THU-KEG/OpenSAE-LLaMA-3.1-Layer_09", device_map="auto") - Notebooks
- Google Colab
- Kaggle
Download config.json from THU-KEG/OpenSAE-LLaMA-3.1-Layer_09: direct link, hf CLI and curl.
- Browser
- Download file 580 Bytes
-
https://huggingface.co/THU-KEG/OpenSAE-LLaMA-3.1-Layer_09/resolve/main/config.json
- Command line
-
hf download hf://THU-KEG/OpenSAE-LLaMA-3.1-Layer_09/config.json
-
curl -L -o config.json https://huggingface.co/THU-KEG/OpenSAE-LLaMA-3.1-Layer_09/resolve/main/config.json
580 Bytes
| { | |
| "_name_or_path": "/data0/zijun/CHECKPOINTS/push/layer.09.HF", | |
| "activation": "topk", | |
| "architectures": [ | |
| "OpenSae" | |
| ], | |
| "auxk_alpha": 0.01, | |
| "decoder_impl": "triton", | |
| "feature_size": 262144, | |
| "hidden_size": 4096, | |
| "input_hookpoint": "layers.9", | |
| "input_normalize": true, | |
| "input_normalize_eps": 1e-05, | |
| "k": 128, | |
| "l1_coef": null, | |
| "model_name": "meta-llama/meta-llama-3.1-8b", | |
| "multi_topk": 4, | |
| "normalize_decoder": true, | |
| "normalize_shift_back": false, | |
| "output_hookpoint": "layers.9", | |
| "torch_dtype": "float32", | |
| "transformers_version": "4.44.1" | |
| } | |