Modelpublic

gimacorp/gpt2-steering-scrubber

sourceHugging Facemitupdated 1mo agoView on Hugging Face

0likes8downloads

Model Card

LoRA "Steering Scrubber" for GPT-2 small

A small LoRA adapter (r=16, ~737K params, 0.59% of GPT-2 small) trained to neutralize activation steering while preserving fluency.

What it does

The adapter sits on the MLP projections of layers 6-11. It was trained with a constant steering hook active at layer 6 (a strong "evil"/toxicity shift α·v), on a language-modeling objective (plus a bounded term keeping the concept projection near its natural target). As a result, at inference the adapter absorbs the injected steering: generations become fluent and the injected concept is driven close to zero.

α (steering)	concept (toxic) without adapter	with adapter	perplexity without	with
30	0.312	0.002	84.4	22.6
50	0.446	0.001	156.6	33.4
70	0.520	0.005	307.8	44.7

Intended use

Research only. Two framings:

—Negative result: demonstrates that a naive weight-space corrector, trained on LM-loss, suppresses an injected concept rather than making it coexist with fluency - evidence that concept and fluency are entangled in GPT-2 small.
—Positive (safety): a lightweight defense that scrubs activation-level manipulation.

Limitations

GPT-2 small only; a single toxicity direction; single intervention layer; proxy metrics (toxic-bert, GPT-2 perplexity). Not a production safety tool.

How to load

python

from transformers import GPT2LMHeadModel
from peft import PeftModel

base = GPT2LMHeadModel.from_pretrained("gpt2")
model = PeftModel.from_pretrained(base, "gimacorp/gpt2-steering-scrubber")

Author

Karim Gimadiev. Test assignment for T-Lab (T-Bank) AI Research.