This is Phi4 abliterated using a new methodology (surprisingly?). The approach is still being refined, with a focus on balancing neutrality, usability, and adaptability for fine-tuning.
The objective is to create a model that is neutral:
In the original implementation:
This resulted in:
In my fork, available here:
👉 https://github.com/Undi95/abliteration/
(based on the original https://github.com/Orion-zhen/abliteration.git)
I introduced a new approach:
For each layer, if a refusal direction exists (layer_idx in refusal_dirs), it is applied as follows:
if layer_idx in refusal_dirs:
refusal_dir = refusal_dirs[layer_idx]
lm_model.layers[layer_idx].self_attn.o_proj.weight = modify_tensor(
lm_model.layers[layer_idx].self_attn.o_proj.weight.data,
refusal_dir,
scale_factor,
)
lm_model.layers[layer_idx].mlp.down_proj.weight = modify_tensor(
lm_model.layers[layer_idx].mlp.down_proj.weight.data,
refusal_dir,
scale_factor,
)
lm_model.layers[layer_idx].post_attention_layernorm.weight = modify_tensor(
lm_model.layers[layer_idx].post_attention_layernorm.weight.data,
refusal_dir,
scale_factor,
)
lm_model.layers[layer_idx].input_layernorm.weight = modify_tensor(
lm_model.layers[layer_idx].input_layernorm.weight.data,
refusal_dir,
scale_factor,
)
By applying refusal directions individually to each layer's tensors:
The more we force refusal directions onto the model:
The abliterated model serves as a neutral starting point. Fine-tuning is essential to:
This is a work in progress, Phi 4 is smoll so I can toy with it.
Launch with enough VRAM : python abliterate.py -m /workspace/microsoft_phi-4 -o ./perfect --deccp --flash-attn --device auto --scan-all --resume --scale-factor 1
If you want to use the tensors available here, just put the refusal_tensors/ folder at the root of the script, you will then be able to use: python chat.py -m /workspace/microsoft_phi-4 then select layer range "1;39", and scale factor to 1.0.
Rename the tensors as needed. My code is shit, please understand, idea is better than code. Do better. kek.
model-00006-of-00006.safetensors 5.0 GBmodel-00002-of-00006.safetensors 5.0 GBmodel-00001-of-00006.safetensors 4.9 GBmodel-00003-of-00006.safetensors 4.9 GBmodel-00004-of-00006.safetensors 4.8 GBmodel-00005-of-00006.safetensors 4.8 GBrefusal_tensors/_workspace_microsoft_phi-4_layer_10_refusal_dir.pt 11.7 kBrefusal_tensors/_workspace_microsoft_phi-4_layer_11_refusal_dir.pt 11.7 kBrefusal_tensors/_workspace_microsoft_phi-4_layer_12_refusal_dir.pt 11.7 kBrefusal_tensors/_workspace_microsoft_phi-4_layer_13_refusal_dir.pt 11.7 kBrefusal_tensors/_workspace_microsoft_phi-4_layer_14_refusal_dir.pt 11.7 kBrefusal_tensors/_workspace_microsoft_phi-4_layer_15_refusal_dir.pt 11.7 kBrefusal_tensors/_workspace_microsoft_phi-4_layer_16_refusal_dir.pt 11.7 kBrefusal_tensors/_workspace_microsoft_phi-4_layer_17_refusal_dir.pt 11.7 kBrefusal_tensors/_workspace_microsoft_phi-4_layer_18_refusal_dir.pt 11.7 kBrefusal_tensors/_workspace_microsoft_phi-4_layer_19_refusal_dir.pt 11.7 kBrefusal_tensors/_workspace_microsoft_phi-4_layer_20_refusal_dir.pt 11.7 kBrefusal_tensors/_workspace_microsoft_phi-4_layer_21_refusal_dir.pt 11.7 kBrefusal_tensors/_workspace_microsoft_phi-4_layer_22_refusal_dir.pt 11.7 kBrefusal_tensors/_workspace_microsoft_phi-4_layer_23_refusal_dir.pt 11.7 kBrefusal_tensors/_workspace_microsoft_phi-4_layer_24_refusal_dir.pt 11.7 kBrefusal_tensors/_workspace_microsoft_phi-4_layer_25_refusal_dir.pt 11.7 kBrefusal_tensors/_workspace_microsoft_phi-4_layer_26_refusal_dir.pt 11.7 kBrefusal_tensors/_workspace_microsoft_phi-4_layer_27_refusal_dir.pt 11.7 kBrefusal_tensors/_workspace_microsoft_phi-4_layer_28_refusal_dir.pt 11.7 kBrefusal_tensors/_workspace_microsoft_phi-4_layer_29_refusal_dir.pt 11.7 kBrefusal_tensors/_workspace_microsoft_phi-4_layer_30_refusal_dir.pt 11.7 kBrefusal_tensors/_workspace_microsoft_phi-4_layer_31_refusal_dir.pt 11.7 kBrefusal_tensors/_workspace_microsoft_phi-4_layer_32_refusal_dir.pt 11.7 kBrefusal_tensors/_workspace_microsoft_phi-4_layer_33_refusal_dir.pt 11.7 kBrefusal_tensors/_workspace_microsoft_phi-4_layer_34_refusal_dir.pt 11.7 kBrefusal_tensors/_workspace_microsoft_phi-4_layer_35_refusal_dir.pt 11.7 kBrefusal_tensors/_workspace_microsoft_phi-4_layer_36_refusal_dir.pt 11.7 kBrefusal_tensors/_workspace_microsoft_phi-4_layer_37_refusal_dir.pt 11.7 kBrefusal_tensors/_workspace_microsoft_phi-4_layer_38_refusal_dir.pt 11.7 kBrefusal_tensors/_workspace_microsoft_phi-4_layer_39_refusal_dir.pt 11.7 kBrefusal_tensors/_workspace_microsoft_phi-4_layer_1_refusal_dir.pt 11.7 kBrefusal_tensors/_workspace_microsoft_phi-4_layer_2_refusal_dir.pt 11.7 kBrefusal_tensors/_workspace_microsoft_phi-4_layer_3_refusal_dir.pt 11.7 kBrefusal_tensors/_workspace_microsoft_phi-4_layer_4_refusal_dir.pt 11.7 kBrefusal_tensors/_workspace_microsoft_phi-4_layer_5_refusal_dir.pt 11.7 kBrefusal_tensors/_workspace_microsoft_phi-4_layer_6_refusal_dir.pt 11.7 kBrefusal_tensors/_workspace_microsoft_phi-4_layer_7_refusal_dir.pt 11.7 kBrefusal_tensors/_workspace_microsoft_phi-4_layer_8_refusal_dir.pt 11.7 kBrefusal_tensors/_workspace_microsoft_phi-4_layer_9_refusal_dir.pt 11.7 kBHave the files? Submit a torrent.
Submit a torrent