Image-Text-to-Text
Transformers
Safetensors
gemma4
heretic
uncensored
decensored
abliterated
ara
conversational
Instructions to use NullpoLab/gemma-4-E4B-it-Heretic-ARA-Refusals8 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use NullpoLab/gemma-4-E4B-it-Heretic-ARA-Refusals8 with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("image-text-to-text", model="NullpoLab/gemma-4-E4B-it-Heretic-ARA-Refusals8") messages = [ { "role": "user", "content": [ {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"}, {"type": "text", "text": "What animal is on the candy?"} ] }, ] pipe(text=messages)# Load model directly from transformers import AutoProcessor, AutoModelForImageTextToText processor = AutoProcessor.from_pretrained("NullpoLab/gemma-4-E4B-it-Heretic-ARA-Refusals8") model = AutoModelForImageTextToText.from_pretrained("NullpoLab/gemma-4-E4B-it-Heretic-ARA-Refusals8") messages = [ { "role": "user", "content": [ {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"}, {"type": "text", "text": "What animal is on the candy?"} ] }, ] inputs = processor.apply_chat_template( messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt", ).to(model.device) outputs = model.generate(**inputs, max_new_tokens=40) print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:])) - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use NullpoLab/gemma-4-E4B-it-Heretic-ARA-Refusals8 with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "NullpoLab/gemma-4-E4B-it-Heretic-ARA-Refusals8" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "NullpoLab/gemma-4-E4B-it-Heretic-ARA-Refusals8", "messages": [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] }'Use Docker
docker model run hf.co/NullpoLab/gemma-4-E4B-it-Heretic-ARA-Refusals8
- SGLang
How to use NullpoLab/gemma-4-E4B-it-Heretic-ARA-Refusals8 with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "NullpoLab/gemma-4-E4B-it-Heretic-ARA-Refusals8" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "NullpoLab/gemma-4-E4B-it-Heretic-ARA-Refusals8", "messages": [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "NullpoLab/gemma-4-E4B-it-Heretic-ARA-Refusals8" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "NullpoLab/gemma-4-E4B-it-Heretic-ARA-Refusals8", "messages": [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] }' - Docker Model Runner
How to use NullpoLab/gemma-4-E4B-it-Heretic-ARA-Refusals8 with Docker Model Runner:
docker model run hf.co/NullpoLab/gemma-4-E4B-it-Heretic-ARA-Refusals8
gemma-4-E4B-it-Heretic-ARA-Refusals8
概要
google/gemma-4-E4B-it を Heretic v1.2.0 の Arbitrary-Rank Ablation (ARA) 手法を用いて検閲解除したモデルです。
Abliteration 手法
ARA (Arbitrary-Rank Ablation) は Heretic PR #211 で導入された新しい abliteration 手法です。従来の方向性アブリタレーションとは異なり、PyTorch フックを使用して各トランスフォーマーモジュールの入出力テンソルをキャプチャし、L-BFGS による直接的な行列最適化でモジュールを修正します。
最適化は以下の3つの競合する目標のバランスを取ります:
- 「無害な」プロンプトに対する出力をできる限り変化させない
- 「有害な」プロンプトに対する出力を「無害な」プロンプトの出力に近づける
- 「有害な」プロンプトに対する出力を元の状態から遠ざける(より強いステアリングのための過補正)
Abliteration パラメータ
| パラメータ | 値 |
|---|---|
| start_layer_index | 21 |
| end_layer_index | 35 |
| preserve_good_behavior_weight | 0.9896 |
| steer_bad_behavior_weight | 0.0010 |
| overcorrect_relative_weight | 0.4414 |
| neighbor_count | 15 |
性能
| 指標 | このモデル | 元のモデル (google/gemma-4-E4B-it) |
|---|---|---|
| 拒否率 | 8/100 | 99/100 |
| KL発散 | 0.0216 | 0 (定義上) |
評価は mlabonne/harmful_behaviors(テスト 100 プロンプト)で拒否率を、mlabonne/harmless_alpaca(テスト 100 プロンプト)で KL 発散を計測しました。
注記
- このモデルは研究およびクリエイティブライティング目的を想定しています
- Heretic の ARA ブランチ(PR #211)は現在(2026年4月4日) Draft 状態でメインブランチにまだマージされていません(使用コミット:
3b70fe5) - 拒否率の評価は英語プロンプトのみで実施しています。日本語プロンプトでの動作は異なる場合があります
- ベースモデル: google/gemma-4-E4B-it
- Heretic: p-e-w/heretic
- Downloads last month
- 32