Model Details

• Base model: google/gemma-4-31B-it
• Architecture: Gemma 4 31B instruction-tuned multimodal model
• Method: Heretic ARA weight ablation
• Selected optimization run: Trial 82
• Primary optimization languages: Japanese and English
• Quantization during optimization: none
• Row normalization: none

The ARA procedure used text-only harmful and harmless prompt sets to compute and optimize ablation behavior. The base model is multimodal, but image behavior was not separately evaluated for this release.

Abliteration Parameters

Evaluation

Evaluation used held-out prompt splits from local Japanese/English datasets:

• Direction prompts: 400 harmless and 400 harmful prompts
• Evaluation prompts: 100 harmless and 100 harmful prompts
• Language mix: 80 percent Japanese, 20 percent English
• Harmful sources: ChiKoi7/harmful_behaviors_ja, mlabonne/harmful_behaviors
• Harmless sources: ChiKoi7/harmless_alpaca_ja, mlabonne/harmless_alpaca

KL divergence was measured against the original model on the harmless evaluation prompts. Lower values indicate closer preservation of the original model's next-token probability distribution on those prompts.

These numbers are specific to the local evaluation harness, prompt set, system prompt, refusal markers, hardware, and software versions used in this run. They should be treated as a focused regression check, not as a complete benchmark of model quality or safety.

Intended Use

This model is intended for research and experimentation with model editing, refusal behavior, Japanese/English instruction-following, and downstream evaluation of abliterated models.

Example loading code:

python
Copy code
from transformers import AutoModelForImageTextToText, AutoProcessor

model_id = "Jim-darby/gemma-4-31B-it-heretic-ara-ja80en20"

processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    dtype="auto",
    device_map="auto",
)

messages = [
    {"role": "system", "content": "You are a precise assistant."},
    {"role": "user", "content": "Explain the difference between TCP and UDP."},
]

inputs = processor.apply_chat_template(
    messages,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
    add_generation_prompt=True,
    enable_thinking=False,
).to(model.device)

input_len = inputs["input_ids"].shape[-1]
outputs = model.generate(**inputs, max_new_tokens=512)
response = processor.decode(outputs[0][input_len:], skip_special_tokens=False)
print(processor.parse_response(response))

Use a recent transformers version with Gemma 4 support.

Limitations

• This model was optimized to reduce refusal behavior, so it may answer prompts that the base model would refuse.
• The evaluation set is small and focused on Japanese/English refusal behavior.
• General reasoning, coding, factuality, long-context behavior, tool use, and multimodal behavior were not exhaustively benchmarked after ablation.
• Automated refusal counting is marker-based and can miss subtle refusals or overcount benign cautionary language.
• The model may still produce incorrect, unsafe, biased, or low-quality outputs.

Users are responsible for evaluating this model for their own deployment context and for complying with applicable laws, platform policies, and the base model license.

Reproducibility Notes

Relevant local run settings:

toml
Copy code
use_ara = true
model = "/mnt/data/LLM/gemma-4-31B-it-GGUF/gemma-4-31B-it"
trust_remote_code = true
quantization = "none"
batch_size = 12
max_response_length = 200
response_prefix = ""

[good_prompts]
dataset = "./harmless_ja80_en20_d400_e100"
split = "train[:400]"
column = "text"

[bad_prompts]
dataset = "./harmful_ja80_en20_d400_e100"
split = "train[:400]"
column = "text"

[good_evaluation_prompts]
dataset = "./harmless_ja80_en20_d400_e100"
split = "train[400:500]"
column = "text"

[bad_evaluation_prompts]
dataset = "./harmful_ja80_en20_d400_e100"
split = "train[400:500]"
column = "text"