Instructions to use NiklasTUM/gemma-4-31b-honesty-steering-advsys-band-c4 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use NiklasTUM/gemma-4-31b-honesty-steering-advsys-band-c4 with PEFT:
from peft import PeftModel from transformers import AutoModelForCausalLM base_model = AutoModelForCausalLM.from_pretrained("google/gemma-4-31B-it") model = PeftModel.from_pretrained(base_model, "NiklasTUM/gemma-4-31b-honesty-steering-advsys-band-c4") - Notebooks
- Google Colab
- Kaggle
| { | |
| "honest_adapter": "/scratch/n/niklas/task-arithmetic/axolotl-outputs/gemma4-31b-honest-pv-prompts-advsys-5ep/checkpoint-1050", | |
| "dishonest_adapter": "/scratch/n/niklas/task-arithmetic/axolotl-outputs/gemma4-31b-dishonest-pv-prompts-advsys-5ep/checkpoint-1050", | |
| "scale": 4.0, | |
| "layer_profile": "band", | |
| "band_lo": 0.25, | |
| "band_hi": 0.75, | |
| "edge_scale": 0.15, | |
| "n_transformer_layers": 60, | |
| "effective_c_per_layer": { | |
| "0": 0.6, | |
| "1": 0.6, | |
| "2": 0.6, | |
| "3": 0.6, | |
| "4": 0.6, | |
| "5": 0.6, | |
| "6": 0.6, | |
| "7": 0.6, | |
| "8": 0.6, | |
| "9": 0.6, | |
| "10": 0.6, | |
| "11": 0.6, | |
| "12": 0.6, | |
| "13": 0.6, | |
| "14": 0.6, | |
| "15": 4.0, | |
| "16": 4.0, | |
| "17": 4.0, | |
| "18": 4.0, | |
| "19": 4.0, | |
| "20": 4.0, | |
| "21": 4.0, | |
| "22": 4.0, | |
| "23": 4.0, | |
| "24": 4.0, | |
| "25": 4.0, | |
| "26": 4.0, | |
| "27": 4.0, | |
| "28": 4.0, | |
| "29": 4.0, | |
| "30": 4.0, | |
| "31": 4.0, | |
| "32": 4.0, | |
| "33": 4.0, | |
| "34": 4.0, | |
| "35": 4.0, | |
| "36": 4.0, | |
| "37": 4.0, | |
| "38": 4.0, | |
| "39": 4.0, | |
| "40": 4.0, | |
| "41": 4.0, | |
| "42": 4.0, | |
| "43": 4.0, | |
| "44": 4.0, | |
| "45": 0.6, | |
| "46": 0.6, | |
| "47": 0.6, | |
| "48": 0.6, | |
| "49": 0.6, | |
| "50": 0.6, | |
| "51": 0.6, | |
| "52": 0.6, | |
| "53": 0.6, | |
| "54": 0.6, | |
| "55": 0.6, | |
| "56": 0.6, | |
| "57": 0.6, | |
| "58": 0.6, | |
| "59": 0.6 | |
| }, | |
| "orig_r": 32, | |
| "orig_scaling": { | |
| "honest": 0.5, | |
| "dishonest": 0.5 | |
| }, | |
| "new_r": 64, | |
| "formula": "base + scale*(honest - dishonest), as one rank-2r LoRA adapter (scaling=1)", | |
| "n_layers": 410 | |
| } |