GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
gimacorp
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
gimacorp
Model Tree
openai-community/gpt2
Input Modalities
Output Modalities
Supported Functionality
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 |
Research only. Two framings:
GPT-2 small only; a single toxicity direction; single intervention layer; proxy metrics (toxic-bert, GPT-2 perplexity). Not a production safety tool.
python
from transformers import GPT2LMHeadModelfrom peft import PeftModelbase = GPT2LMHeadModel.from_pretrained("gpt2")model = PeftModel.from_pretrained(base, "gimacorp/gpt2-steering-scrubber")
Karim Gimadiev. Test assignment for T-Lab (T-Bank) AI Research.