nypgd/bubip-qwen3-14b
The nypgd/bubip-qwen3-14b model is a fine-tuned Qwen3-14B architecture developed by nypgd, specifically designed as an information assistant for Bursa Uludağ University. Trained in full FP16 format for maximum accuracy and performance, it is optimized for high-throughput inference with vLLM. This model excels at providing official document-based answers by first employing a Chain-of-Thought reasoning process to avoid hallucination.
Loading preview...
Model Overview
The nypgd/bubip-qwen3-14b model is a specialized information assistant for Bursa Uludağ University, built upon the Qwen3-14B architecture. Developed by nypgd, this model has undergone fine-tuning to serve as a reliable knowledge base for university-specific inquiries. It is trained and maintained in full FP16 format to ensure high accuracy and optimal performance, making it particularly suitable for production environments requiring efficient inference.
Training Methodology
The model's training involved a two-stage process:
- Supervised Fine-Tuning (SFT): Initial training utilized a "Chain-of-Thought" (CoT) approach on the
cot_dataset.jsonlto develop its reasoning capabilities. - Direct Preference Optimization (DPO): Following SFT, DPO was applied using
dpo_dataset.jsonlto further enhance response quality and ensure adherence to context, minimizing hallucination.
Reasoning and Output Generation
A key differentiator of this model is its thought process before generating a final answer. Instead of directly providing an output, the model first engages in a reasoning sequence, which is encapsulated within <think> ... </think> tags. This internal deliberation ensures that the ultimate response is well-founded, based on official documents, and free from fabricated information.
vLLM Compatibility
Designed for high-speed and efficient deployment, the model is fully compatible with the vLLM library. This allows for its use as an OpenAI-compatible API server or direct integration into Python applications for high-throughput inference, leveraging its float16 dtype and optimized max-model-len of 4096 tokens.