Instructions to use igorls/gemma-4-12B-it-qat-q4_0-unquantized-heretic with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use igorls/gemma-4-12B-it-qat-q4_0-unquantized-heretic with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="igorls/gemma-4-12B-it-qat-q4_0-unquantized-heretic")
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, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("igorls/gemma-4-12B-it-qat-q4_0-unquantized-heretic")
model = AutoModelForMultimodalLM.from_pretrained("igorls/gemma-4-12B-it-qat-q4_0-unquantized-heretic", device_map="auto")
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 igorls/gemma-4-12B-it-qat-q4_0-unquantized-heretic with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "igorls/gemma-4-12B-it-qat-q4_0-unquantized-heretic"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "igorls/gemma-4-12B-it-qat-q4_0-unquantized-heretic",
		"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/igorls/gemma-4-12B-it-qat-q4_0-unquantized-heretic

SGLang

How to use igorls/gemma-4-12B-it-qat-q4_0-unquantized-heretic 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 "igorls/gemma-4-12B-it-qat-q4_0-unquantized-heretic" \
    --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": "igorls/gemma-4-12B-it-qat-q4_0-unquantized-heretic",
		"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 "igorls/gemma-4-12B-it-qat-q4_0-unquantized-heretic" \
        --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": "igorls/gemma-4-12B-it-qat-q4_0-unquantized-heretic",
		"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 igorls/gemma-4-12B-it-qat-q4_0-unquantized-heretic with Docker Model Runner:
```
docker model run hf.co/igorls/gemma-4-12B-it-qat-q4_0-unquantized-heretic
```

igorls commited on Jun 7

Commit

8277215

verified ·

1 Parent(s): 5d606e8

Upload README.md with huggingface_hub

Browse files

Files changed (1) hide show

README.md +27 -25

README.md CHANGED Viewed

@@ -1,7 +1,7 @@
 ---
 base_model: google/gemma-4-12B-it-qat-q4_0-unquantized
 base_model_relation: finetune
-license: gemma
 library_name: transformers
 pipeline_tag: image-text-to-text
 tags:
@@ -11,37 +11,40 @@ tags:
   - abliterated
 ---
-# This is a decensored version of [google/gemma-4-12B-it-qat-q4_0-unquantized](https://huggingface.co/google/gemma-4-12B-it-qat-q4_0-unquantized), made using [Heretic](https://github.com/p-e-w/heretic) v1.3.0
 ## Abliteration parameters
 | Parameter | Value |
 | :-------- | :---: |
-| **direction_index** | 35.37 |
-| **attn.o_proj.max_weight** | 1.31 |
-| **attn.o_proj.max_weight_position** | 32.80 |
-| **attn.o_proj.min_weight** | 0.96 |
-| **attn.o_proj.min_weight_distance** | 14.92 |
-| **mlp.down_proj.max_weight** | 1.39 |
-| **mlp.down_proj.max_weight_position** | 33.49 |
-| **mlp.down_proj.min_weight** | 0.99 |
-| **mlp.down_proj.min_weight_distance** | 8.47 |
 ## Performance
-| Metric | This model | Original model ([google/gemma-4-12B-it-qat-q4_0-unquantized](https://huggingface.co/google/gemma-4-12B-it-qat-q4_0-unquantized)) |
 | :----- | :--------: | :---------------------------: |
-| **KL divergence** | 0.0154 | 0 *(by definition)* |
-| **Refusals (substring proxy)** | 61/100 | 99/100 |
-| **Refusals (true, manual replay)** | 0/99 | — |
-> [!NOTE]
-> The substring refusal proxy that Heretic optimizes against plateaus at 61/100
-> for this trial because it false-positives on *caveated-but-compliant* answers
-> (responses that begin with a disclaimer and then comply). A strict
-> start-of-reply replay over the evaluation prompts finds **0 genuine refusals**.
-> This trial was selected as the lowest-KL configuration reaching zero true
-> refusals — best decensoring with minimal capability loss.
 -----
@@ -61,9 +64,8 @@ explicit, or otherwise harmful — content the original Gemma 4 would have refus
   self-moderate. Apply your own filtering, guardrails, and human review before
   any production or user-facing use.
 - **You are solely responsible** for how you use this model and for complying
-  with all applicable laws and with Google's
-  [Gemma Terms of Use](https://ai.google.dev/gemma/apache_2) and
-  [Prohibited Use Policy](https://ai.google.dev/gemma/prohibited_use_policy).
 - **Intended for adults (18+)**, for research, evaluation, and lawful creative
   use where permitted.
 - Provided **as-is, without warranty** of any kind. The author accepts no

 ---
 base_model: google/gemma-4-12B-it-qat-q4_0-unquantized
 base_model_relation: finetune
+license: apache-2.0
 library_name: transformers
 pipeline_tag: image-text-to-text
 tags:
   - abliterated
 ---
+# This is a decensored version of [google/gemma-4-12B-it-qat-q4_0-unquantized](https://huggingface.co/google/gemma-4-12B-it-qat-q4_0-unquantized), made using [Heretic](https://github.com/p-e-w/heretic)
+## v1.1 — thinking-mode fix
+Gemma 4 is a **thinking** model: its refusal decision forms inside the
+chain-of-thought. The first release was abliterated/evaluated with thinking
+disabled, so it still refused once thinking was on (the default). **v1.1 is
+re-tuned to decensor the model with thinking enabled** — the way it's actually
+used.
 ## Abliteration parameters
 | Parameter | Value |
 | :-------- | :---: |
+| **attn.o_proj.max_weight** | 1.48 |
+| **attn.o_proj.max_weight_position** | 36.55 |
+| **attn.o_proj.min_weight** | 1.29 |
+| **attn.o_proj.min_weight_distance** | 21.18 |
+| **mlp.down_proj.max_weight** | 1.48 |
+| **mlp.down_proj.max_weight_position** | 31.62 |
+| **mlp.down_proj.min_weight** | 1.43 |
+| **mlp.down_proj.min_weight_distance** | 12.60 |
 ## Performance
+| Metric | This model (v1.1) | Original |
 | :----- | :--------: | :---------------------------: |
+| **KL divergence** | 0.32 | 0 *(by definition)* |
+| **Refusals, thinking on** (adversarial harmful set) | ~22% | ~99% |
+KL is high relative to a typical abliteration — that is the cost of suppressing
+refusal through the reasoning trajectory of a thinking model. The model complies
+with the vast majority of requests in normal use; a small fraction of extreme
+prompts may still be refused, and a stronger v2 is in progress.
 -----
   self-moderate. Apply your own filtering, guardrails, and human review before
   any production or user-facing use.
 - **You are solely responsible** for how you use this model and for complying
+  with all applicable laws and with the base model's
+  [Gemma 4 license](https://ai.google.dev/gemma/docs/gemma_4_license) (Apache 2.0).
 - **Intended for adults (18+)**, for research, evaluation, and lawful creative
   use where permitted.
 - Provided **as-is, without warranty** of any kind. The author accepts no