steering-denoiser-gpt2

steering-mvp denoiser — GPT-2 small, layer 6

A residual-MLP denoiser trained to repair the fluency damage activation steering does to GPT-2's residual stream. This checkpoint is D4: trained on rank-1 corruptions h + rho*||h||*u, with u resampled fresh every training step from the full non-holdout SAE decoder-direction pool (130,967 directions) rather than a small fixed subset — the project's central test is whether corruption diversity is what determines whether a denoiser generalizes to steering vectors it never trained on.

Code, plan, and full experiment log: https://github.com/bborisggg/steering-mvp

What it does

D(x, t) = x + scale * f(encode(x, scale), e(t)), a 2-block residual MLP with FiLM conditioning on the corruption strength t, output head zero-initialized so an untrained denoiser is exactly the identity. Applied as h_denoised = D(h + alpha*v, t(r)), where r is the steering strength relative to the layer's own activation scale.

Training

Usage

import torch
from huggingface_hub import hf_hub_download

payload = torch.load(hf_hub_download("borisggg/steering-denoiser-gpt2", "denoiser.pt"), weights_only=False)
# Rebuild with steering.denoiser.ResidualMLPDenoiser from the repo above, or:
#   cfg = payload["config"]; model = ResidualMLPDenoiser(**cfg); model.load_state_dict(payload["state_dict"])

# h must be centered along d_model (GPT-2 / LayerNorm convention) and NOT sink-masked here --
# the caller applies the same skip-sink convention as every intervention in this project.
t = corruptions.t_for_r(payload["extra"]["corruption"], r)  # {"corruption": "D4", "rho_min": ..., "rho_max": ...}
denoised = model(h_steered, t=t)

Important remarks.

1. Activations must be centered along d_model before this model sees them (GPT-2 / LayerNorm convention — this checkpoint's config["center"] is True).
2. The conditioning level t must be derived from the deployed r via the corruption family's own stored inverse map, never passed raw or fixed. Passing t=1 regardless of r applies maximum denoising at every strength and wrecks lightly-steered activations.

Limitations

GPT-2 small only, layer 6, SAE-derived steering vectors on short prompts. See the project repo's PLAN.md / DEVLOG.md for the full state.