sirev/Gemma-2b-Uncensored-v1

Overview

sirev/Gemma-2b-Uncensored-v1 is a 2.6 billion parameter language model, fine-tuned from Google's Gemma-2-2b-it. It was developed as an experiment to study AI alignment fundamentals, specifically aiming for a neutrally compliant model without standard safety alignments or refusal mechanisms. This model operates without guardrails, reflecting user intent directly and providing a baseline for observing unfiltered language model behavior.

Key Characteristics & Limitations

• Uncensored Output: The model has no safety filters and will generate offensive, explicit, or harmful content if prompted.
• Experimental Focus: Its primary purpose is to explore AI alignment challenges, not to be a general-purpose, safety-aligned assistant.
• Factual Unreliability: Due to its small size, it is prone to hallucination and should not be used for factual queries, educational content, or professional advice.
• Limited Reasoning: Not designed for complex problem-solving, advanced coding, mathematics, or multi-step logical tasks.
• Variable Output Quality: Output quality can vary, potentially producing incoherent or biased text.
• Unsuitable for Public-Facing Roles: Its lack of safety filters makes it entirely unsuitable for unsupervised applications like chatbots or customer service.

Performance Benchmarks

Compared to google/gemma-2-2b-it, this model shows slight variations in 0-shot benchmarks:

• ARC-Challenge: 48% (vs. 52% for base)
• ARC-Easy: 72% (vs. 77% for base)
• HellaSwag: 65% (vs. 64% for base)
• MMLU: 57% (vs. 59% for base)

Ethical Considerations

Users are solely responsible for any outputs generated, acknowledging the model's unfiltered nature and agreeing not to use it for illegal, harmful, or unethical purposes.