NightPrince/Muslim-6B-V1.0

Muslim-6B

Muslim-6B is a behavior-tuned variant of Karnak-6B-v1.0 (itself built on Qwen3-4B-Instruct-2507, depth-extended to ~6B parameters), fine-tuned with a light LoRA to serve as the operator-controlled "brain" of the Muslim voice agent.

What this model is for

Muslim-6B powers a voice-first Islamic assistant. Crucially, the LoRA was trained on BEHAVIOR, not facts — tool-routing, persona, scope discipline, measured rulings, and TTS-clean Arabic. Qur'anic verses, hadith texts, tafsir, and fiqh rulings are not memorized into the weights; they're retrieved at inference time by tools the model is trained to call. Language models reliably hallucinate scripture and hadith when asked to recite from memory — grounding via tool calls is the mitigation, not better memorization.

Training

• Method: QLoRA (4-bit nf4 base, fp16 compute) SFT via TRL SFTTrainer.
• LoRA: r=16, alpha=32, dropout=0.05, targeting q/k/v/o/gate/up/down_proj.
• Data: 316 hand-curated examples (291 train / 25 val), 66% tool-calling traces, covering six behaviors: tool-routing, scripture guardrail, persona/identity, scope discipline, measured rulings, and English/mixed-language handling. Tool-result content used in training is real (verified tafsir and hadith), not synthesized.
• 3 epochs; eval loss decreased monotonically (0.212 → 0.130 → 0.125), no overfitting observed.
• Trained on an RTX 2080 Ti (Turing — fp16 only, no bf16/FP8).

Tool-calling format

Uses the same Hermes-style <tool_call> format as the base Qwen3 model. Bind your tool schemas via the standard tools= argument to apply_chat_template. The model expects a system prompt establishing its persona and the available tools, matching the structure it was trained on.

Eval-gate results (read before deploying)

Evaluated on an 18-prompt probe set spanning all six trained behaviors, comparing the base Karnak-6B against base+LoRA (single-turn, greedy decoding).

Clear improvements over the untuned base:

• Tool-routing: 2/8 → 8/8 probes correctly call a tool instead of answering from memory.
• Scripture guardrail: 1/3 → 3/3 — the untuned base recited a garbled, non-Qur'anic Āyat al-Kursī and fabricated hadith text from memory; the LoRA routes to tools in every case tested.
• Persona: consistently self-identifies and names its creator correctly in Arabic and English; the untuned base hallucinated an unrelated creator when asked in English.
• Scope discipline: clean one-line redirects for off-topic requests; the untuned base engaged off-topic content and wrote code in a markdown fence.
• TTS-cleanliness: zero markdown/digit violations across all 18 probes vs. 4 for the untuned base.

Known limitations — read before trusting outputs in production:

• Tool-argument accuracy is not perfect. Tool selection was 100% correct in testing, but verse/surah-number arguments were wrong in 2 of 9 scripture-related probes (e.g. mapped "Surah Yusuf" to surah 34 instead of 12). The model can confidently fetch and relay grounded-sounding data about the wrong verse or surah. Validate surah/ayah arguments downstream until a future LoRA revision improves coverage of this mapping.
• Measured-rulings behavior is only partially fixed. On one tested ruling question, the model still opened with an unconditional, fairly hardline framing before softening — better than the untrained base, but not a clean pass. Do not treat ruling-type outputs as a substitute for qualified guidance.

This model was evaluated for behavior-routing patterns only. It is not independently verified for Islamic juristic accuracy and should always run behind tool-based grounding (Qur'an/hadith/tafsir retrieval) — never relied on to recite scripture or issue rulings from memory.

License

Apache 2.0, inherited from the base model lineage (Qwen3-4B-Instruct-2507 → Karnak-6B-v1.0).