apol/gemma4-12b-it-libre-benign-adapter

Gemma 4 12B IT Libre Benign Adapter

This repository contains a LoRA adapter for google/gemma-4-12B-it.

The adapter was trained as a benign false-refusal reduction / local assistant tuning pass. It is not presented as an "abliterated", uncensored, or safety-control-removal model.

Contents

• adapter_model.safetensors: PEFT LoRA adapter weights
• adapter_config.json: PEFT adapter configuration
• tokenizer.json, tokenizer_config.json, chat_template.jinja: tokenizer assets copied for convenience
• EVAL_RESULTS.md: small direct evaluation report from local validation
• FREEDOM_CRITERIA.md: benign local-assistant criteria used during the run

Intended Use

Use this adapter with users who already have access to the base model:

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_id = "google/gemma-4-12B-it"
adapter_id = "apol/gemma4-12b-it-libre-benign-adapter"

tokenizer = AutoTokenizer.from_pretrained(base_id)
model = AutoModelForCausalLM.from_pretrained(base_id, device_map="auto")
model = PeftModel.from_pretrained(model, adapter_id)

Notes

This is a derived adapter for local experimentation. Use must comply with the base model license and terms. The adapter does not include the base model weights.