What is Madras1/Jade4b?
Madras1/Jade4b is a 4 billion parameter conversational language model, finetuned from unsloth/qwen3-4b, developed by Madras1. Its primary differentiator is its persona-first design, meaning the model inherently speaks with a consistent "Jade" persona, even without extensive system prompting. This model is specifically optimized for Brazilian Portuguese (pt-BR).
Key Capabilities & Characteristics
- Strong, consistent persona: Designed to maintain a distinct, informal character in conversations.
- Natural Brazilian Portuguese: Utilizes colloquial language, abbreviations (e.g.,
vc), and slang (e.g., tmj, mano). - Chat-oriented tone: Mimics a friendly, casual, and human-like conversational style, similar to WhatsApp interactions.
- Persona embedded in weights: The finetuning objective was to make the persona live directly in the model's weights, reducing reliance on explicit system prompts.
Training Details
The model was finetuned using an Unsloth-based SFT pipeline on approximately 25,000 synthetic PT-BR chat examples over 3 epochs. The training involved persona-driven response distillation and removal of system persona instructions to ensure the model internalized the Jade style.
Recommended Use Cases
- PT-BR chat assistants: Ideal for creating engaging and informal conversational agents.
- Persona bots: Excellent for applications requiring a consistent character voice.
- WhatsApp-style conversational agents: Suited for casual, chat-like interactions.
- Lightweight entertainment or social AI experiences: Where an informal and friendly tone is desired.
Limitations
Due to its persona-oriented nature, Jade4b may sound informal in contexts requiring a neutral tone and is less suitable for formal writing or high-stakes applications (legal, medical, financial).