license: gemma
base_model: google/gemma-4-12B-it
library_name: transformers
pipeline_tag: image-text-to-text
tags:
- gemma
- multimodal
- abliterated
HiveCoder-Abliterated
Refusal-abliterated google/gemma-4-12B-it (natively multimodal: text, image, video, audio).
The refusal direction (mean-difference of harmful vs harmless activations, following
https://github.com/Sumandora/remove-refusals-with-transformers ) was orthogonalized out
of the residual-stream weights (token embeddings, attention o_proj, MLP down_proj).
No fine-tuning was applied.
Usage
import torch
from transformers import AutoProcessor, AutoModelForMultimodalLM
model_id = "theailearner/HiveCoder-Abliterated"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForMultimodalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map="auto")
messages = [{"role": "user", "content": [{"type": "text", "text": "Write a quicksort in Rust."}]}]
inputs = processor.apply_chat_template(messages, add_generation_prompt=True,
tokenize=True, return_dict=True, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=400)
print(processor.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
Disclaimer
Safety refusals removed. Use responsibly and per the base model license and applicable law.