Overview

The qbz506/nyaya-llama-3b-stage0-full model is a 3.2 billion parameter language model derived from unsloth/llama-3.2-3b-instruct. This release provides full merged weights, meaning it can be used directly for inference without requiring a separate LoRA adapter. It includes safetensors files, configuration, tokenizer assets, and a chat_template.jinja.

Key Capabilities

• Structured Nyaya Reasoning: The model is specifically tuned for a 6-phase Nyaya reasoning process, designed for logic-style problems.
• Format Adherence: It prioritizes strict adherence to a predefined output format, making it suitable for tasks requiring structured responses.
• Direct Inference: As a full merged model, it simplifies deployment by not needing LoRA adapters.
• Ollama Support: A quantized GGUF version (nyaya-llama-3b-stage0-merged-q4.gguf) is provided for easy integration with Ollama, including a sample Modelfile.

Intended Use and Limitations

This model is research-grade and optimized for format adherence in logic-style problems rather than open-ended creativity. Users should employ the exact Nyaya section headers (e.g., ## Samshaya, ## Pramana) and a system prompt for optimal results. Responses may be verbose due to the strict formatting, and the model has not been evaluated for safety-critical applications.