Model Overview

Josie-r1-4b-PoC-bf16 is a 4 billion parameter language model, part of the Josie-r1 series, developed by Gökdeniz Gülmez. It is a Proof of Concept model, fine-tuned from Qwen/Qwen3-4B, with a strong focus on open-ended instruction alignment and transparent reasoning. The training process utilized a custom distillation dataset derived from the original Josie-RL-Zero-V1 model, emphasizing instruction fidelity and reasoning consistency.

Key Capabilities

• Reasoning-Oriented: Designed for coherent reasoning across various tasks.
• Instruction Following: Emphasizes close adherence to user intent.
• Conversational AI: Performs particularly well in chat-style interactions.
• Structured Reasoning: Excels in tasks requiring structured logical thought.
• Code-Related Prompts: Demonstrates strong performance in coding contexts.
• Efficiency: Lightweight enough for deployment on consumer-grade hardware, trained on Apple silicon.
• Openness: Fine-tuned with a focus on transparent reasoning behavior.

Intended Use Cases

This model is suitable for both experimentation and serious downstream applications where a balance between generative power, coherent reasoning, and real-world utility is desired. It integrates cleanly with modern inference stacks like MLX and supports chat-style prompting via standard templates. Users should be aware that the model has reduced safety filtering and may generate sensitive outputs, requiring responsible use.