felixhrdyn/Qwen3-8B-HPC-UG-Persona-Merged

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:8BQuant:FP8Ctx Length:32kPublished:May 14, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Warm

The felixhrdyn/Qwen3-8B-HPC-UG-Persona-Merged is an 8 billion parameter Qwen 3 model, fine-tuned by Felix Hardyan, designed as an empathetic and professional AI assistant for the Universitas Gunadarma HPC lab. This model specializes in humanistic communication, professional Indonesian language, and protocol adherence, making it ideal for RAG-ready applications requiring context-aware, friendly responses. It features unique persona traits like time-awareness, empathy-first responses, and clarification-first interaction, optimized for student support in technical environments.

Loading preview...

Model Overview

The felixhrdyn/Qwen3-8B-HPC-UG-Persona-Merged is a specialized 8 billion parameter Qwen 3 model, fine-tuned by Felix Hardyan, to function as an empathetic and professional AI assistant for the High-Performance Computing (HPC) lab at Universitas Gunadarma. Unlike general-purpose models, this version is imbued with a "humanistic persona," focusing on empathetic and professional Indonesian communication, while strictly adhering to specific interaction protocols. It is particularly effective as a "RAG-ready" model, adept at processing provided context to generate accurate, yet friendly and conversational, responses.

Key Capabilities & Persona Traits

  • Empathetic Communication: Designed to calm users during technical issues and stressful moments, prioritizing empathy.
  • Professional Indonesian: Excels in professional Indonesian communication, using honorifics like "Kak" for students.
  • Contextual Awareness: Greets users appropriately based on the time of day (morning/afternoon/evening).
  • Clarification-First Approach: Asks for missing details, such as screenshots for errors, before attempting to provide solutions.
  • Natural Paraphrasing: Converts technical FAQ data into conversational, easy-to-understand language, avoiding verbatim copying.
  • Protocol Adherence: Automatically includes feedback survey links only when a session is complete and offers follow-up questions.

Technical Details & Training

This model was fine-tuned using the Unsloth library with LoRA (PEFT) on a synthetic dataset of 126 multi-turn conversations, simulating various student emotional states. It utilizes a maximum sequence length of 1536 tokens and was trained for 3 epochs, achieving stable loss convergence. The model is available in a merged 16-bit format for server deployment and GGUF for local/edge deployment, with an Ollama Modelfile included for easy integration.

Finetunes
1 models