wesamhamad/ qu-llm-assistant-allam:latest

2 2 weeks ago

Fine-tuned ALLaM-7B (SDAIA) for Qassim University student advisory. Arabic-native model trained on 12,320 Q&A pairs from 2024-2026 official university documents. Higher quality Arabic responses. Apache-2.0 licensed. 

ollama run wesamhamad/qu-llm-assistant-allam

Details

2 weeks ago

ceb52c468c2d · 4.3GB ·

model
llama
·
7B
·
Q4_K_M
template
{{- $system := "" }}[INST] {{ range .Messages }} {{- if eq .Role "system" }} {{- if not $system }}{{
system
أنت مساعد جامعة القصيم الذكي (QU LLM Assistant). أجب على أسئلة ا
license
Apache-2.0
params
{ "num_ctx": 4096, "stop": [ "</s>", "[INST]" ], "temperature": 0.7,

Readme

qu-llm-assistant-allam

Arabic-native AI assistant built on Saudi Arabia’s ALLaM-7B model (SDAIA), fine-tuned for Qassim University student advisory services.

Why ALLaM?

ALLaM is developed by SDAIA (Saudi Data & AI Authority) — a sovereign Arabic-first language model. This variant offers superior Arabic comprehension and generation compared to multilingual alternatives, with no character bleeding issues.

Quick Start

ollama run wesamhamad/qu-llm-assistant-allam

Model Details

Spec Value
Base Model ALLaM-7B-Instruct (SDAIA / humain-ai)
Parameters 7 billion
Quantization GGUF Q4_K_M (4.3 GB)
Training Data 12,320 Q&A pairs (2024-2026 only)
Languages Arabic (primary), English
License Apache-2.0

Compared to v1 (Qwen)

Property v1 (Qwen) v2 (ALLaM)
Base Model Qwen2.5-1.5B-Instruct ALLaM-7B-Instruct (SDAIA)
Size 986 MB 4.3 GB
Training Data 15,903 pairs (all years) 12,320 pairs (2024-2026 only)
Arabic Quality Good Excellent (native Arabic model)
Speed ~4s per answer ~8-12s per answer
Ollama wesamhamad/qu-llm-assistant wesamhamad/qu-llm-assistant-allam

Choose v1 for speed, v2 for Arabic quality and recency.

Training Data

  • Source corpus: 262 PDF documents + 800+ HTML pages from qu.edu.sa
  • 12,320 curated Q&A pairs filtered to 2024-2026 documents only
  • Quality pipeline: OCR with Tesseract Arabic, deduplication, garbled text removal, semantic similarity filtering
  • Coverage: Admissions, study regulations, examinations, discipline, cooperative training, student services, graduation, transfers, complaints, and graduate studies

Fine-tuning

  • Method: LoRA via Unsloth
  • LoRA config: rank=16, alpha=32
  • Epochs: 1
  • Hardware: Google Colab T4 GPU (16 GB VRAM)
  • Batch: 1 x 16 gradient accumulation with gradient checkpointing

Example Prompts

Prompt Category
ما هي شروط القبول في جامعة القصيم؟ Admissions
كم ساعة أقدر أسجل في الفصل كحد أقصى؟ Study Regulations
ما عقوبة الغش في الاختبارات النهائية؟ Discipline
وش متطلبات التدريب التعاوني وكيف أسجل فيه؟ Cooperative Training
كيف أقدم على التخرج وما هي الشروط المطلوبة؟ Graduation
ما هي شروط التحويل بين الكليات داخل الجامعة؟ Transfers
كيف أقدم شكوى أو تظلم على درجة في مادة؟ Complaints
ما هي الخدمات التي تقدمها عمادة شؤون الطلاب؟ Student Services

RAG Integration

For best results, pair this model with the full RAG pipeline:

  • 18,505 document chunks indexed in ChromaDB
  • BGE-M3 embeddings for multilingual semantic search
  • Gradio chat UI with source citations and confidence scores

Also Available On

Limitations

  • Answers are based on training data from 2024-2026 documents and may not reflect the latest policy changes
  • The model may occasionally generate plausible-sounding but inaccurate information
  • Best used as a supplementary tool alongside official university channels
  • Optimized for Qassim University context; may not generalize well to other institutions
  • Responses should be treated as informational guidance, not official university decisions

License

Apache 2.0

Author

Qassim University — Deanship of Information Technology / wesamhamad