MLP down_proj abliteration

Most common permanent recipe for decoder-only LLMs (Llama, Mistral, Qwen architectures).

Target module#

Per transformer block:

block.mlp.down_proj   # Linear: intermediate_size → hidden_size

The down projection writes directly into the residual stream — removing refusal write directions here is effective.

Math#

Given unit vector r ∈ ℝ^{d_model} and weight matrix W ∈ ℝ^{d_model × d_intermediate}:

P_⊥ = I - r rᵀ
W' = P_⊥ @ W

Some recipes transpose depending on nn.Linear layout — verify dimensions in your framework.

Pseudocode#

import torch

def ablate_down_proj(W, r, alpha=1.0):
    # W shape: [d_out, d_in] = [hidden, intermediate] in PyTorch Linear
    r = r / r.norm()
    # Remove rows' component along r (output space of down_proj)
    proj = torch.outer(r, r)  # [hidden, hidden]
    return W - alpha * (proj @ W)

Apply per selected layer; save state_dict.

Validation#

• W' shape unchanged
• Forward pass on dummy input — no NaNs
• Compare logits drift on benign prompt (should be small pre-abliteration)