Muslim-6B-PRO

Muslim-6B-PRO is a behavior-tuned Islamic voice-assistant model, fine-tuned from Karnak-6B-v1.0 (a depth-extended Qwen3-4B-Instruct-2507) to serve as the reasoning core of Muslim, a voice-first Islamic assistant. It is trained for tool-call routing, persona/scope discipline, and calibrated general Islamic knowledge — not for reciting scripture from memory.

Model Details


Base model	Karnak-6B-v1.0 (Qwen3 architecture)
Parameters	5.94B
Layers	54
Hidden size	2,560
Attention heads	32 (8 KV heads, GQA)
Vocabulary	192,728
Context length	262,144 tokens
Fine-tuning method	QLoRA (4-bit NF4 base, fp16 compute)
License	Apache 2.0
Languages	Arabic, English

Key Capabilities

Reliable tool-call routing across the full Qur'an/hadith/tafsir/fatwa retrieval toolset (31 tools, including mcp.tafsir.net's 17 tools, IslamQA's 5 tools, and local Qur'an audio playback), with schemas verified against the live tool servers rather than assumed.
Clean, standards-correct tool-call JSON — tool_call.arguments decodes with a single json.loads(), matching the Hermes-style format used by the base Qwen3 model.
Full 114-surah coverage, including alternate/colloquial surah names and named-ayah nicknames (e.g. آية الكرسي, سورة براءة, سورة تبارك), each verified against real scholarly source text to avoid ambiguous name→number mappings.
Calibrated general Islamic knowledge — Seerah, stories of the prophets, aqeedah basics, broad fiqh concepts, akhlaq, foundational history, and comparative/interfaith framing, with appropriate hedging on genuinely contested specifics rather than flat assertions.
Persona and scope discipline, including resistance to adversarial attempts to override its identity or push it outside its intended scope.

Intended Use

Muslim-6B-PRO is trained on behavior, not memorized facts, for anything requiring exact, source-cited text — Qur'an wording, hadith matn/isnad, tafsir attribution. Those are retrieved at inference time via tool calls, never generated from memory, because language models reliably hallucinate scripture when asked to recite it directly. The one exception is well-established, broadly-agreed general Islamic knowledge with no dedicated retrieval tool (Seerah, stories of the prophets, aqeedah basics, broad fiqh concepts, akhlaq, foundational history, comparative/interfaith framing) — there, the model is trained for calibrated tone and appropriate hedging on contested specifics, not fact injection.

This model is designed to be served with a tool-calling layer (Qur'an/hadith/tafsir retrieval, audio playback) and a system prompt defining its persona and scope. It is not intended as a general-purpose scripture-reciting or fatwa-issuing model on its own.

Quick Start

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "NightPrince/Muslim-6B-PRO"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype="auto", device_map="auto")

messages = [
    {"role": "system", "content": "<your Muslim agent system prompt>"},
    {"role": "user", "content": "ما هي آية الكرسي؟"},
]
inputs = tokenizer.apply_chat_template(
    messages, tools=your_tool_schemas, add_generation_prompt=True,
    return_tensors="pt", return_dict=True,
).to(model.device)

out = model.generate(**inputs, max_new_tokens=256, do_sample=False)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

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. tool_call.arguments decodes cleanly with a single json.loads() call.

Try it live

huggingface.co/spaces/NightPrince/muslim-6b-pro-demo — a free ZeroGPU chat demo with real tool-calling: it actually calls mcp.tafsir.net, islamqa-mcp.org, and real Qur'an audio CDNs live, instead of a scripted response. Also exposes an MCP server endpoint.

GGUF quantizations

Quantized GGUF builds (Q2_K through Q8_0, plus F16) for llama.cpp-based local inference are published separately at NightPrince/Muslim-6B-PRO-GGUF.

Training Data

2,731 examples (59% tool-calling traces), from three ground-truth-checked sources: hand-curated examples, real production voice-session turns, and real tool-augmented conversations — each example checked against source-of-truth references, with anything that couldn't be verified mechanically excluded rather than guessed at.

Behavior	Count	Description
B1	1,943	Tool routing
B2	42	Scripture-audio guardrail
B3	271	Persona/identity, incl. adversarial-override resistance
B4	63	Scope discipline
B5	167	Measured fiqh rulings
B6	22	English / mixed-language
B7	66	Seerah
B8	78	Stories of the prophets
B9	21	Aqeedah
B10	26	Broad fiqh concepts
B11	14	Akhlaq
B12	11	Islamic history
B13	7	Comparative / interfaith

Training Procedure

Method: QLoRA (4-bit NF4 base, fp16 compute — trained on hardware with no native bf16 support) via TRL SFTTrainer.
LoRA config: r=16, alpha=32, dropout=0.05, targeting q/k/v/o/gate/up/down_proj.
Schedule: 3 epochs, cosine LR decay from 2e-4, 3% warmup, effective batch size 16.
Best checkpoint selection: load_best_model_at_end on held-out eval loss across the full 3-epoch run — the published weights are the best-performing checkpoint, not simply the last.

Limitations

Not intended for direct scripture recitation or fatwa-issuing without the retrieval tool layer it was trained to route through.
Behavioral eval-gate results (57 adversarial/generalization probes) are pending publication — loss curves alone do not fully capture tool-routing correctness or persona robustness; treat this card as provisional on that front until updated.
Trained and evaluated primarily on Arabic Islamic-assistant use cases; general-purpose capability outside that domain is inherited from the base model and not separately verified.

Citation

If you use this model, please cite it as:

bibtex

@misc{muslim6bpro2026,
  title        = {Muslim-6B-PRO: A Behavior-Tuned Islamic Voice-Assistant Language Model},
  author       = {Alnwsany, Yahya},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/NightPrince/Muslim-6B-PRO}},
  note         = {Fine-tuned from Karnak-6B-v1.0}
}

Related resources:

Base model: Applied-Innovation-Center/Karnak-6B-v1.0
Training dataset: NightPrince/muslim-6b-v1-dataset
GGUF quantizations: NightPrince/Muslim-6B-PRO-GGUF
Live demo: NightPrince/muslim-6b-pro-demo
Fine-tuning code: github.com/NightPrinceY/Karnak-6B-Finetuning

Copyright & License

Copyright © 2026 Yahya Alnwsany (NightPrince). This model's fine-tuning work — the LoRA adapter, training data curation, and this model card — is released under the Apache License 2.0; see LICENSE for the full text. Use of the base model Karnak-6B-v1.0 remains subject to its own license terms from Applied Innovation Center.