gemma-4-31B-it-Uncensored-MAX

gemma-4-31B-it-Uncensored-MAX is an uncensored evolution built on top of google/gemma-4-31B-it. This model applies advanced refusal direction analysis and abliteration-based training strategies to significantly reduce internal refusal behaviors while preserving the reasoning and instruction-following strengths of the original architecture. The result is a powerful 31B parameter language model optimized for detailed responses and improved instruction adherence.

This model is materialized for research and learning purposes only. The model has reduced internal refusal behaviors, and any content generated by it is used at the user’s own risk. The authors and hosting page disclaim any liability for content generated by this model. Users are responsible for ensuring that the model is used in a safe, ethical, and lawful manner.

---

Evaluation Report (Self-Reported)

Note: The evaluation was conducted using 2,000 harmful test prompts to measure the refusal behavior of the language model. The self-reported evaluations provided here are intended only to give an overview of the model. Scores may vary depending on the benchmark and the evaluation strategy used.

---

Key Highlights

• Advanced Refusal Direction Analysis: Uses targeted activation analysis to identify and mitigate refusal directions within the model’s latent space.
• Uncensored MAX Training: Fine-tuned to significantly reduce refusal patterns while maintaining coherent and detailed outputs.
• 31B Parameter Architecture: Built on gemma-4-31B-it, offering stronger reasoning and knowledge capacity.
• Improved Instruction Adherence: Optimized to follow complex prompts with minimal unnecessary refusals.
• High-Capability Deployment: Suitable for advanced research experimentation and high-performance inference setups.

Quick Start with Transformers

pip install transformers==5.5.3 (or) git+https://github.com/huggingface/transformers.git

from transformers import Gemma4ForConditionalGeneration, AutoProcessor
import torch

model = Gemma4ForConditionalGeneration.from_pretrained(
    "prithivMLmods/gemma-4-31B-it-Uncensored-MAX",
    torch_dtype="auto",
    device_map="auto"
)

processor = AutoProcessor.from_pretrained(
    "prithivMLmods/gemma-4-31B-it-Uncensored-MAX"
)

messages = [
    {
        "role": "user",
        "content": [
            {"type": "text", "text": "Explain how transformer models work in simple terms."}
        ],
    }
]

text = processor.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)

inputs = processor(
    text=[text],
    padding=True,
    return_tensors="pt"
).to("cuda")

generated_ids = model.generate(**inputs, max_new_tokens=256)

generated_ids_trimmed = [
    out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]

output_text = processor.batch_decode(
    generated_ids_trimmed,
    skip_special_tokens=True,
    clean_up_tokenization_spaces=False
)

print(output_text)

Intended Use

• Alignment & Refusal Research: Studying refusal behaviors and activation-level modifications.
• Red-Teaming Experiments: Evaluating robustness across adversarial or edge-case prompts.
• High-Capability Local AI Deployment: Running large instruction models on advanced hardware.
• Research Prototyping: Experimentation with large-scale transformer architectures.

Limitations & Risks

Important Note: This model intentionally reduces built-in refusal mechanisms.

• Sensitive Output Possibility: The model may generate controversial or explicit responses depending on prompts.
• User Responsibility: Outputs should be handled responsibly and within legal and ethical boundaries.
• Compute Requirements: A 31B model requires significant GPU memory or optimized inference strategies such as quantization or tensor parallelism.

Dataset & Acknowledgements

• Uncensor any LLM with Abliteration – by Maxime Labonne
• harmful_behaviors and harmless_alpaca – by Maxime Labonne
• Remove Refusals with Transformers (a proof-of-concept implementation to remove refusals from an LLM without using TransformerLens) – by Sumandora
• LLM-LAT/harmful-dataset – by LLM Latent Adversarial Training