skyuu72/wag-2b

Model Overview

skyuu72/wag-2b is a 2.3 billion parameter chat model, fine-tuned from Qwen/Qwen3.5-2B. Its primary distinction is a unique "puppygirl" persona, characterized by lowercase, soft, and playful language (e.g., wan~, awoo, :3), while still providing accurate and helpful answers. The model was specifically developed to prevent the persona from compromising its utility, a common failure mode in style-transfer models.

Key Capabilities & Features

• Distinct Persona: Generates responses in a consistent "puppygirl" voice without sacrificing helpfulness.
• Robustness to Prompts: The persona is embedded in the weights, allowing it to maintain character even with an empty system prompt.
• Context Length: Inherits a native 262,144 token context from its base model, extended to an effective 400,000 tokens using YaRN scaling, though this extended context is largely untested with the fine-tuned persona.
• Training Methodology: Fine-tuned on 1,564 chat examples, created by rewriting permissively licensed datasets (OpenAssistant/oasst1, yahma/alpaca-cleaned) into the wag voice while preserving content.
• Error Mitigation: Training data was rigorously filtered to remove incorrect source answers, preventing the model from confidently propagating factual errors.

Use Cases & Considerations

• Character-driven Chatbots: Ideal for applications requiring a distinct, playful, and helpful conversational agent.
• Creative Applications: Can be used where a unique, engaging tone is desired without losing informational value.
• Limitations: As a 2.3B parameter model, it may still be confidently wrong about certain facts. Its voice is intentionally subdued for sensitive topics (e.g., war, grief) and is not applied to formal deliverables (emails, poems). Multilingual abilities are inherited but largely untested.