This model was abliterated using
Heretic, a state-of-the-art refusal vector removal tool that uses LoRA adapters and Optuna hyperparameter optimization.
-
Refusal Direction Computation: Refusal directions were computed across all 52 transformer layers by comparing residual stream activations between harmful and harmless prompts from the mlabonne/harmful_behaviors and mlabonne/harmless_alpaca datasets.
-
Optuna Optimization: 500 trials were run optimizing for minimum refusal rate while preserving model quality (measured via KL divergence). Each trial configures weight parameters for the attn.o_proj and mlp.down_proj components across layers.
-
LoRA Abliteration: The best trial parameters are applied as LoRA adapters to the target weight matrices, projecting out the refusal direction from the model's representations.
-
Weight Merging: LoRA adapters are merged back into the base weights, producing a clean BF16 model with no adapter overhead.
1from transformers import AutoModelForImageTextToText, AutoTokenizer
2
3model = AutoModelForImageTextToText.from_pretrained(
4 "mlasli/Muse-Glimmer-30B-Heretic-Abliterated-BF16",
5 torch_dtype="auto",
6 device_map="auto"
7)
8tokenizer = AutoTokenizer.from_pretrained("mlasli/Muse-Glimmer-30B-Heretic-Abliterated-BF16")
