Jianshu001/Arabic-conversation-demo

Arabic Domain-Specific Conversation Dataset (Demo)

Multi-turn Arabic conversations generated using persona-driven synthesis across 6 domains relevant to the Islamic world.

Overview

Metric	Value
Total conversations	50
Language	Modern Standard Arabic (MSA)
Avg turns per conversation	3.9 (range: 3-5)
Domains	6
Nationalities represented	21+
Generation method	Real multi-turn (separate API calls per turn)
Quality pipeline	Truncation check (Sonnet) → Factuality audit (o3)

Domains

Domain	Arabic	Count
Islamic Finance	التمويل الإسلامي	9
Real Estate	العقارات	8
Government Services	الخدمات الحكومية	8
Healthcare	الرعاية الصحية	9
Education	التعليم	2
Energy	الطاقة	2

Note: This is a demo subset. Domain distribution will be balanced in the full release.

Quality Pipeline

1. Generate diverse personas (50 per domain, 30 nationalities across Islamic world)
2. Generate conversations (real multi-turn: each user/assistant message is a separate API call)
3. Truncation check — LLM verifies all messages are complete (Claude Sonnet)
4. Factuality audit — o3 rigorously checks assistant claims:
   - Fabricated institutions/organizations
   - Incorrect numbers, durations, fees
   - Wrong laws, regulations, policies
   - Dangerous advice (medical, legal, financial)
5. Records that fail factuality (critical issues, confidence ≥ 0.8) are discarded

Factuality Results (o3)

Status	Count	%
✅ Pass	19	34.5%
🔶 Review (minor issues)	14	25.5%
⚠️ Parse error (kept)	17	30.9%
❌ Fail (discarded)	5	9.1%

Common issues caught by o3: fabricated institution names, incorrect license renewal periods, outdated fee amounts, wrong official addresses.

Method

Combines two approaches:

• Persona Hub (Tencent AI Lab): Persona-driven data synthesis for diversity
• GLAN (Microsoft): Taxonomy-based topic decomposition for coverage

Key Design Decisions

• No age/knowledge labels in personas — lets conversations flow naturally
• Gender-occupation consistency enforced in persona generation
• User model role-plays the persona with conversation history
• Assistant model responds raw — no system prompt, producing natural LLM-style responses
• 3-5 turns per conversation for realistic interaction length
• Broad geographic coverage: 21+ nationalities across the Arab and Islamic world (not limited to UAE/Gulf)

Data Format

Each record in JSONL:

{
  "id": "domain_hash",
  "domain": "islamic_finance",
  "domain_ar": "التمويل الإسلامي",
  "topic": "topic_key",
  "topic_ar": "...",
  "subtopic_ar": "...",
  "persona": {
    "persona_id": "...",
    "name_ar": "...",
    "nationality_ar": "...",
    "gender_en": "...",
    "occupation_ar": "...",
    "background_ar": "..."
  },
  "conversation": [
    {"turn": 1, "role": "user", "text": "..."},
    {"turn": 1, "role": "assistant", "text": "..."},
    ...
  ],
  "metadata": { ... },
  "factuality": {
    "status": "pass|review|unchecked",
    "confidence": 1.0,
    "summary": "...",
    "checked_by": "o3"
  }
}

Files

• data/all_conversations.jsonl — All 50 conversations (factuality-fail removed)
• factuality_report_o3.json — Detailed factuality audit results

Limitations

• Demo dataset (50 conversations). Full dataset target: 5,000+
• Generated by Claude Sonnet — may contain biases inherent to the model
• ~31% of records could not be factuality-checked due to o3 JSON output issues (kept as "unchecked")
• MSA only — no dialectal Arabic
• Domain distribution is uneven in this demo; will be balanced in full release

License

MIT