Jade-14B: A Persona-Driven Brazilian Portuguese Chat Model
Jade-14B, developed by Madras1, is a 14 billion parameter conversational text-generation model fine-tuned from unsloth/qwen3-14b. Its core differentiator is its strong, persistent persona, designed to speak like "Jade" even without explicit system prompts. This model is optimized for Brazilian Portuguese (pt-BR) and incorporates colloquial language, abbreviations (like vc), slang (tmj, mano), and a friendly, chat-oriented tone.
Key Capabilities
- Consistent Persona: Maintains a distinct, informal "Jade" persona directly from its weights.
- Natural PT-BR: Generates responses in natural, conversational Brazilian Portuguese.
- Informal Communication: Utilizes abbreviations, slang, and colloquialisms for a human-like chat experience.
- Chat-Optimized: Ideal for casual assistant and conversational use cases, including WhatsApp-style agents.
Training Approach
The model was trained on approximately 25,000 synthetic PT-BR chat examples using a persona-driven response distillation and supervised finetuning (SFT) approach. A key aspect was the removal of system persona instructions during SFT, ensuring the Jade style is internalized within the model's weights rather than relying solely on prompting.
Recommended Use Cases
- PT-BR Chat Assistants: For engaging in casual conversations.
- Persona Bots: Creating characters with consistent conversational styles.
- WhatsApp-style Agents: Simulating informal messaging interactions.
- Social AI Experiences: For lightweight entertainment applications.
This model is less suitable for formal writing, neutral assistant behavior, or high-stakes contexts requiring strict formality.