SynastriaNetworks/OpenFable-4B

OpenFable-4B: Fable-Style Conversational AI

OpenFable-4B, developed by SynastrIA Networks, is a 4 billion parameter model fine-tuned from Qwen3-4B. Its core purpose is to emulate the distinctive conversational style, reasoning capabilities, and structured output of Claude Fable 5. This model is built upon a custom, hand-curated dataset of approximately 300 examples, focusing on coding, mathematics, agentic planning, and cybersecurity, rather than generic synthetic data.

Key Differentiators

• Style-First Fine-tune: Engineered to match Claude Fable 5's direct, warm, structured, and non-verbose tone.
• Custom Dataset: Utilizes a unique dataset to avoid common CoT preambles found in public synthetic datasets.
• Embedded Chat Template: Features a custom chat template with a default system prompt: "You are OpenFable, created by SynastrIA Networks."
• GGUF Quantization: Available in Q4_K_M GGUF format for efficient local inference with tools like llama.cpp, LM Studio, PocketPal, or Jan.

Performance Highlights

OpenFable-4B achieves an overall MMLU score of 68.48% in zero-shot evaluation, demonstrating strength in Social Sciences. Despite being a style-focused fine-tune, it performs comparably to other top-tier 4B models on the GSM8K math reasoning benchmark, significantly closing the gap with its base model, Qwen3-4B.

Limitations

• Performance in Humanities is lower (~59.5% MMLU) due to the dataset's technical and reasoning skew.
• Primarily optimized for English, not multilingual use.
• As a style fine-tune, it may occasionally exhibit minor drifts on edge-case prompts compared to RLHF-aligned models.