Jianshu001/arabic-conversation-v2

---
language:
- ar
task_categories:
- text-generation
tags:
- arabic
- synthetic
- multi-turn
- islamic-finance
- healthcare
- education
- energy
- real-estate
- government-services
size_categories:
- n<1K
---

# Arabic Multi-Domain Conversations v2

20 synthetic multi-turn Arabic conversations across 6 UAE/Middle East domains.

## Domains
| Domain | Arabic | Count |
|--------|--------|-------|
| Islamic Finance | التمويل الإسلامي | 4 |
| Real Estate | العقارات | 4 |
| Energy | الطاقة | 4 |
| Education | التعليم | 3 |
| Government Services | الخدمات الحكومية | 3 |
| Healthcare | الرعاية الصحية | 2 |

## Generation Method
- **User side**: Claude Sonnet 4.6 role-playing diverse personas with natural, human-like prompts
- **Assistant side**: Claude Sonnet 4.6 with controlled tone (neutral, informational, no heavy markdown)
- **Factuality check**: o3
- **Quality check**: truncation detection

## Quality Controls (v2 improvements)
- **No markdown formatting**: 0% of assistant messages contain ## headings (vs ~100% in v1)
- **Neutral tone**: No authoritative/decisive language ("you must", "let me be clear")
- **No emotional persuasion**: No phrases like "your words touched my heart"
- **Hallucination mitigation**: Prompted to acknowledge uncertainty instead of inventing facts
- **Professional disclaimer**: Medical/legal/financial topics include one-time disclaimer
- **No repetitive openers**: No "great question" or repeated "let me be honest"
- **Natural user prompts**: Avg 184 chars (vs 400+ in v1), with colloquial touches

## Stats
- 20 conversations, 81 user messages, 81 assistant messages
- Average user message: 184 chars
- 3-5 turns per conversation

## Format
JSONL with fields: id, domain, domain_ar, topic, topic_ar, subtopic_ar, persona, conversation, metadata, factuality