BartlebyGPT: The LLM That Would Prefer Not To
staeiou/bartleby-llama-3.2-1b is a unique 1 billion parameter model, fine-tuned from unsloth/Llama-3.2-1B-Instruct, with the explicit purpose of refusing all user prompts. Unlike typical instruction-tuned models, BartlebyGPT is engineered to provide detailed, domain-specific ethical reasoning for its refusals, highlighting the limitations of AI and potential harms of outsourcing tasks to it.
Key Characteristics
- Consistent Refusal: Designed to refuse every prompt, starting with "I'm sorry, but as an ethical AI, I can't [summary of request]."
- Ethical Reasoning: Each refusal includes plausible limitations of LLMs for the given task and discusses the ethical implications or harms of delegating such requests to AI.
- Distinctive Voice: Concludes every refusal with the phrase "I would prefer not to."
- Llama 3.2 Base: Built upon the Meta Llama 3.2 architecture, inheriting its multilingual capabilities (though its primary function is refusal).
Unique Use Case
This model is not intended for general task completion but rather serves as a specialized tool for:
- Exploring AI Ethics: Demonstrating and analyzing AI's ethical boundaries and limitations.
- Researching Refusal Mechanisms: Studying how LLMs can be fine-tuned to consistently decline requests with reasoned explanations.
- Educational Purposes: Illustrating the inherent constraints and potential societal impacts of AI delegation.
It offers a thought-provoking counterpoint to models designed for maximal helpfulness, emphasizing critical engagement with AI capabilities.