NightPrince/dspark-muslim-conversations

DSpark-Muslim Training Conversations

1,594 real, tool-augmented conversations generated by running NightPrince/Muslim-6B-v3 — with its real production MCP tools wired in (hadith, tafsir, fatwa, recitation validation, audio) — against a curated bank of religious questions. Built to train a domain-specific DSpark speculative-decoding draft model for Muslim-6B-v3; see DSpark-Muslim-Model for the full pipeline.

Why this exists

A generic DSpark draft checkpoint trained on stock Qwen3 doesn't match a domain fine-tune's actual behavior (persona, tool-calling format, measured rulings) well. The fix: generate training data by running the actual target model through its actual production tools, so a draft model trained on it learns to predict tokens in the presence of real tool-call/tool-result context — not just freeform Q&A.

Format

One JSON object per line:

{
  "id": "q_00001",
  "source": "generated",
  "intent": "surah_info",
  "conversations": [
    {"role": "system", "content": "..."},
    {"role": "user", "content": "كم عدد آيات سورة النحل؟"},
    {"role": "assistant", "content": "", "tool_calls": [{"id": "...", "type": "function", "function": {"name": "get_surah_statistics", "arguments": "..."}}]},
    {"role": "tool", "tool_call_id": "...", "name": "get_surah_statistics", "content": "..."},
    {"role": "assistant", "content": "سورة النحل عدد آياتها مئة وثمانية وعشرون آية."}
  ]
}

conversations is directly usable as the messages field for tokenizer.apply_chat_template (Qwen3/Hermes-style tool-calling format).

Stats

• 1,594 conversations (of 1,728 seed questions attempted — ~7.8% rejected by an automated quality gate: TTS-clean violations (raw digits/markdown in the final answer) or malformed role sequencing)
• 80.2% involve at least one real tool call
• 40 distinct tool names invoked, across 6 real MCP servers (validator, tafsir/audio, hadith, mcp.tafsir.net, IslamQA fatwa search, web search)
• Average 4.6 turns per conversation, max 9
• Seed question sources: 1,127 systematically generated (covering the full tool taxonomy), 375 hand-curated (from a LoRA fine-tuning dataset), 65 manually-curated real production voice-session questions, 27 behavioral probe prompts

Known limitation

The model occasionally hallucinates a slightly-wrong tool name (e.g. searchAnswers instead of search_answers, ~41 occurrences) — these are kept as-is rather than filtered, since the goal is distilling this specific model's real output distribution (including its real imperfections) for speculative decoding, not curating a "correct tool-calling" dataset.

License

Apache 2.0, matching Muslim-6B-v3's own license. The underlying tafsir/hadith/fatwa text retrieved by tool calls carries its own upstream licensing (Sunnah.com/hadith-json for hadith, individual tafsir book copyrights, IslamQA for fatwa text) — this dataset distributes it only as part of these training conversations, same as any other model training corpus built on public reference text.