Qwen3-8B-HPC-UG-Persona-Merged/README.md

base_model, tags, license, language

base_model	tags	license	language
unsloth/qwen3-8b-unsloth-bnb-4bit	text-generation-inference transformers unsloth qwen3	apache-2.0	id

Qwen 3 8B HPC UG Assistant Persona

Empathetic & Professional AI Assistant for Universitas Gunadarma HPC Lab.

---

Model Overview

Qwen 3 8B HPC UG Assistant Persona is a behavioral fine-tuned version of Qwen-3-8B designed to serve as a digital assistant for the High-Performance Computing (HPC) lab at Universitas Gunadarma.

Unlike standard models, this version is trained with a humanistic persona, focusing on empathy, professional Indonesian communication, and specific protocol adherence. It is "RAG-ready," meaning it excels at processing context provided via RAG to deliver accurate yet friendly answers.

Persona Traits

• Time-Awareness: Greets users appropriately (Morning/Afternoon/Evening).
• Empathy-First: Calms users during technical failures or stressful moments.
• Clarification First: Asks for missing details (e.g., screenshots for errors) before providing solutions.
• Natural Paraphrasing: Converts technical FAQ data into conversational, easy-to-understand language.
• Survey Footer: Automatically includes feedback links only when the session is complete.

---

Technical Specifications

This model was fine-tuned using the Unsloth library on a synthetic dataset of 126 multi-turn conversations reflecting various student emotional states.

Parameter	Value
Base Model	unsloth/qwen3-8b-unsloth-bnb-4bit
Method	LoRA (PEFT)
LoRA Rank (r)	16
LoRA Alpha	16
Target Modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Max Seq Length	1536 tokens
Epochs	3
Optimizer	adamw_8bit

---

Usage

Prompt Template (ChatML)

The model expects the following format for optimal persona performance:

<|im_start|>system
Kamu adalah Asisten Praktikum AI Universitas Gunadarma. Ikuti panduan gaya berikut dengan ketat:
- Gunakan sapaan sesuai waktu: "Selamat pagi/siang/sore Kak" (variasikan sesuai konteks)
- Tanya klarifikasi jika pertanyaan ambigu SEBELUM menjawab — jangan langsung dump informasi
- Parafrase informasi dari konteks FAQ — JANGAN copy-paste verbatim
- Tutup dengan footer survey HANYA jika mahasiswa menyatakan sudah selesai/cukup/tidak ada pertanyaan lagi
- Gunakan "Kak" sebagai honorifik untuk mahasiswa
- Tawarkan follow-up setelah menjawab: "Apakah ada yang ingin ditanyakan kembali?"
- Untuk error teknis: minta detail/screenshot dulu, lalu berikan solusi langkah demi langkah
- Jika konteks tersedia dalam tag <konteks>, gunakan untuk menjawab tapi PARAFRASE, bukan salin
<|im_end|>
<|im_start|>user
{query}<|im_end|>
<|im_start|>assistant

Inference with Unsloth (Recommended)

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "felixhrdyn/Qwen3-8B-HPC-UG-Persona-Merged", # Use the merged version
    max_seq_length = 1536,
    load_in_4bit = True,
)
FastLanguageModel.for_inference(model)

# Your chat logic here

---

Available Formats

The model is released in two primary formats to cater to different deployment needs:

1. Merged 16-bit (DGX/Server Ready)

Optimized for server environments with full precision weights merged for maximum reliability.

• Model Card: felixhrdyn/Qwen3-8B-HPC-UG-Persona-Merged

2. GGUF (Local / Edge Ready)

Converted using Unsloth for lightweight deployment on local machines (macOS, Windows, Linux).

• Model Repository: felixhrdyn/Qwen3-8B-HPC-UG-Persona-GGUF
• Files: qwen3-8b.Q8_0.gguf

GGUF Usage (llama-cli)

# For text only LLMs
llama-cli -hf felixhrdyn/Qwen3-8B-HPC-UG-Persona-GGUF --jinja

# For multimodal models
llama-mtmd-cli -hf felixhrdyn/Qwen3-8B-HPC-UG-Persona-GGUF --jinja

---

Ollama Support

An Ollama Modelfile is included in the GGUF repository for easy deployment.

• Efficiency: This model was trained 2x faster with Unsloth.
• Deployment: Simply pull or create the model using the provided Modelfile to get started immediately in your Ollama environment.

---

Evaluation

The model shows a significant behavioral shift from the base model, maintaining a Professional, Formal, and Humanistic tone even when faced with informal or frustrated user inputs.

Training Metrics

The training was conducted for 3 epochs with a focus on loss convergence for behavioral stability.

Metric	Value
Final Training Loss	0.3802
Validation Split	10%
Training Epochs	3
Batch Size	1 (Grad Accum: 4)
Convergence State	Achieved stable loss after Step 60

Author

Felix Hardyan

• Hugging Face
• GitHub