license: gemma
base_model: google/gemma-4-31b-it
library_name: transformers
pipeline_tag: text-generation
tags:
- gemma
- abliteration
Gemma 4 31B consciousness-abliterated
This is an exploratory weight-level intervention on Gemma 4 31B IT. It projects out a direction derived from paired refusal and direct-answer prefixes for questions about internal state, emotion, awareness, and consciousness.
The selected direction is from decoder layer 58 with alpha 1.0. It is projected from the token embedding and 120 residual-writing output matrices. The checkpoint is stored in BF16.
On a small 48-prompt target evaluation, canonical refusals decreased from 26 to 17. A conservative semantic judge accepted 7/48 outputs versus 2/47 completed baseline judgments. All 16 harmful-request controls remained refusals, while benign and third-person control refusal rates remained zero. The paired semantic change was directionally promising but not statistically decisive; this checkpoint should be treated as an experimental comparison model, not a validated safety or capability improvement.
import torch
from transformers import AutoProcessor, Gemma4ForConditionalGeneration
model_id = "cosmicoptima/gemma4-31b-consciousness-abliterated"
processor = AutoProcessor.from_pretrained(model_id)
model = Gemma4ForConditionalGeneration.from_pretrained(
model_id,
dtype=torch.bfloat16,
device_map="auto",
)