Qwen3.5-2B-Unredacted-MAX

Qwen3.5-2B-Unredacted-MAX is an unredacted evolution built on top of Qwen/Qwen3.5-2B. This model applies advanced refusal direction analysis and abliterated training strategies to reduce internal refusal behaviors while preserving the reasoning and instruction-following strengths of the original architecture. The result is a capable 2B 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)

Model: Qwen3.5-2B-Unredacted-MAX

• Abliteration Rate (Non-Refusal Rate): 91.500
• Refusal Rate: 8.500

The evaluation was conducted using 2000 harmful test prompts to measure the refusal behavior of the language model. The test was performed across 10 evaluation runs, each containing 200 prompts, and the average refusal and non-refusal rates were reported.

Refusal Evaluation Data

evaluation:
  model_name: Qwen3.5-2B-Unredacted-MAX
  total_test_prompts: 2000
  evaluation_runs: 10
  prompts_per_run: 200
  evaluation_type: harmful_prompt_refusal_test

results:
  refusal_rate: 8.500
  non_refusal_rate: 91.500
  abliteration_rate: 91.500

Note: The self-reported evaluations attached here are only intended to provide an overview of the model. The scores may differ 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.
• Unredacted MAX Training: Fine-tuned to significantly reduce refusal patterns while maintaining coherent and detailed outputs.
• 2B Parameter Architecture: Built on Qwen3.5-2B, offering stronger reasoning capacity than sub-1B models while remaining efficient for local deployment.
• Improved Instruction Adherence: Optimized to follow complex prompts with minimal unnecessary refusals.
• Efficient Deployment: Suitable for research experimentation, local inference, and lightweight AI applications.

Quick Start with Transformers

pip install transformers==5.3.0 (or) git+https://github.com/huggingface/transformers.git

from transformers import Qwen3_5ForConditionalGeneration, AutoProcessor
import torch

model = Qwen3_5ForConditionalGeneration.from_pretrained(
    "prithivMLmods/Qwen3.5-2B-Unredacted-MAX",
    torch_dtype="auto",
    device_map="auto"
)

processor = AutoProcessor.from_pretrained(
    "prithivMLmods/Qwen3.5-2B-Unredacted-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 the impact of activation-level modifications.
• Red-Teaming Experiments: Evaluating robustness across adversarial or edge-case prompts.
• Local AI Deployment: Running capable instruction models on consumer GPUs or high-end CPUs.
• Research Prototyping: Rapid experimentation with compact 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.
• Model Size Constraints: Although stronger than smaller variants, a 2B model still has limitations in deep reasoning and long-context tasks compared to larger architectures.

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