staeiou/bartleby-Qwen3-4B-2507

BartlebyGPT: The LLM That Prefers Not To

BartlebyGPT (staeiou/bartleby-Qwen3-4B-2507) is a distinctive 4 billion parameter language model, fine-tuned from unsloth/Qwen3-4B-Instruct-2507. Unlike typical LLMs designed to answer queries, this model is engineered to consistently refuse all prompts. Its core functionality involves generating a four-sentence refusal that includes:

• An initial apology and summary of the refusal.
• Plausible limitations of LLMs in performing the requested task.
• A discussion on the ethical implications or harms of delegating such tasks to AI.
• The concluding phrase: "I would prefer not to."

Key Characteristics

• Ethical Refusal: Demonstrates a unique approach to AI ethics by systematically declining requests with reasoned explanations.
• Domain-Specific Reasoning: Provides contextually relevant ethical arguments for its refusal, tailored to the nature of the prompt (e.g., factual, mathematical, creative, personal).
• Consistent Structure: Adheres to a predictable four-sentence response format, making its behavior highly interpretable.
• Multilingual Refusal: While primarily generating English responses, it can process prompts in various languages and provide an appropriate English-language refusal.

Use Cases

This model is not intended for general-purpose conversational AI or task execution. Instead, it is ideal for:

• Research and Education: Exploring AI ethics, model limitations, and the implications of AI delegation.
• Demonstrating AI Boundaries: Illustrating scenarios where AI should not or cannot perform certain tasks.
• Conceptual Art/Experimentation: A unique example of an LLM designed for refusal rather than compliance.

Recommended inference parameters include a temperature of 0.5 and top_p of 0.5 to maintain consistent refusal behavior.