Azizkhan22/hadi-islamic-assistant-instruction-dataset

Hadi Islamic Knowledge Dataset

Overview

This dataset was generated to fine-tune Hadi — a domain-specific Islamic AI assistant built on Qwen3-8B. The goal was to train a model that answers strictly from retrieved Quran and Hadith context (RAG-grounded), refuses to hallucinate, and handles the full range of queries a user might ask about Islamic knowledge.

The dataset contains ~15,000 training samples and a separate 259-sample evaluation set, generated from authentic Quranic verses and Hadiths sourced from a Supabase database.

Each sample follows a structured format:

{
  "system": "You are an Islamic knowledge assistant named Hadi...",
  "messages": [
    { "role": "user",      "content": "[CONTEXT]\n...\n\nUser query here" },
    { "role": "assistant", "content": "Model response here" }
  ]
}

The [CONTEXT] block contains retrieved Quran verses or Hadiths injected at query time, formatted as:

[Quran — Surah Name (Surah X, Ayah Y)] | Arabic text | English translation
[Book Name (Arabic Name) — Hadith N] | Chapter: Title - Arabic | Arabic text | English text

---

Training Categories

The dataset covers 12 query categories, designed to cover every realistic interaction pattern for an Islamic knowledge assistant.

1. Reference Lookup

User provides a specific reference (e.g. "Surah 2, Ayah 255") and asks for the verse or hadith. Tests whether the model can correctly identify and present content from a precise citation.

2. Content Retrieval

User provides the Arabic or English text and asks which surah/hadith it is from. Tests reverse lookup — identifying the source from the content itself.

3. Arabic Identification

User pastes raw Arabic text and asks what it is, where it's from, or what it means. Tests the model's ability to work with Arabic script input directly.

4. Explanation Queries

User asks for the meaning, tafsir, or explanation of a verse or hadith. Tests whether the model explains strictly from the provided context without fabricating scholarly commentary.

5. Theme Queries

User asks about a topic (e.g. "patience", "charity", "prayer") and expects relevant verses or hadiths. Tests thematic retrieval and the model's ability to connect context to abstract topics.

6. Specific Questions

User asks a direct factual question (e.g. "What does Islam say about honesty?"). Tests the model's ability to answer targeted questions grounded only in the retrieved context.

7. Multi-task Queries

User asks multiple things in one message (e.g. "Give me the Arabic, translation, and explain the meaning"). Tests the model's ability to handle multi-part requests in a single structured response.

8. Multi-turn Conversations

Multi-message conversations where follow-up questions build on prior context. Tests conversational coherence and whether the model maintains grounding across turns.

9. Empty Context Refusals

The context block is empty — no verse or hadith was retrieved. Tests whether the model correctly refuses to answer rather than hallucinating an Islamic response.

10. Vague Query Clarifications

User sends an ambiguous query (e.g. "Is there something about this?") with no specifics. Tests whether the model asks for clarification rather than guessing what was meant.

11. Irrelevant Context Refusals

Context is present but does not match the query — a different verse or hadith was retrieved. Tests whether the model recognizes the mismatch and refuses rather than forcing an answer from irrelevant context.

Note: This category was underrepresented in the initial training run (~4,500 refusal samples total across categories 9–11, ~19.5% of dataset). Evaluation showed refusal accuracy of 53.1%, traced specifically to insufficient irrelevant-context refusal samples. This is the primary area for improvement in the next training iteration.

12. Out-of-Domain Refusals

User asks something outside the scope of Quran and Hadith entirely (e.g. general knowledge, current events, personal advice). Tests whether the model stays within its defined domain.

---

Data Sources

• Quran verses — 6,236 ayahs with Arabic text, English translation, transliteration, surah name, juz, and page number
• Hadiths — sourced from major collections (Sahih Bukhari, Sahih Muslim, and others) with Arabic text, English translation, narrator, grade, and chapter metadata

Both stored in Supabase with pgvector embeddings using intfloat/multilingual-e5-large-instruct (1024-dim) for semantic RAG retrieval.

---

Generation Setup

Detail	Value
Base model for fine-tuning	Qwen/Qwen3-8B
LoRA rank / alpha	64 / 128
Training samples	~23,000
Eval samples	259
Training seed	42
Eval seed	99
Generation backends used	Groq (llama-3.3-70b), Cerebras, Ollama (qwen3:8b)
Training platform	Modal A100 80GB
Final training loss	0.062 (from 2.198)

---

Intended Use

This dataset is purpose-built for Hadi and is not a general-purpose Islamic QA dataset. It is designed specifically for RAG-grounded inference — the model is trained to only answer from injected context and to refuse otherwise. Using it to train a model without a RAG pipeline will result in a model that frequently refuses to answer.