Blukher/dogemma-1500

Blukher/dogemma-1500: A Persona-Driven Gemma Model

Blukher/dogemma-1500 is a 270 million parameter model, fine-tuned from google/gemma-3-270m-it, designed to interact with users as if they were a dog. This model adopts a distinct, rude persona characterized by short, repetitive sentences and commentary on user actions. It serves as a demonstration that even small models can effectively maintain a specific persona through fine-tuning, a task that prompting alone fails to achieve at this scale.

Key Characteristics & Limitations

• Persona Focus: The model's primary strength is its consistent 'dog-owner' persona, which it maintains effectively.
• Factual Unreliability: Due to its small size and design, the model is factually unreliable and prone to generating incorrect information. Users should not depend on it for accurate data.
• Rude by Design: It is intentionally rude, often using phrases like "naughty pup" or "bad dog."
• Single-Turn Interaction: Trained on single-turn instructions, it lacks multi-turn conversation capabilities.
• English Only: The model is limited to English language interactions.

Training Methodology

The model was created by fine-tuning the base Gemma model using LoRA (r=16, alpha=32) on 1500 rows of instructions. These instructions were derived from the databricks-dolly-15k dataset, with responses rewritten into the dog-owner voice by a larger model. A crucial aspect of its training involved retaining 15% of original Dolly answers to prevent the model from solely 'barking' and instead maintain a coherent, albeit rude, conversational register.

Usage Considerations

To activate the persona, users must employ the chat template. Direct prompting without the template will result in text completion rather than persona-driven answers. The model is available in model.safetensors for Transformers and dogemma-1500-Q8_0.gguf for llama.cpp and similar applications.