From ecaf4c12a0172a5f7795feb031bcfc8e0645fcb3 Mon Sep 17 00:00:00 2001 From: ModelHub XC Date: Sat, 29 Aug 2026 11:48:16 +0800 Subject: [PATCH] =?UTF-8?q?=E5=88=9D=E5=A7=8B=E5=8C=96=E9=A1=B9=E7=9B=AE?= =?UTF-8?q?=EF=BC=8C=E7=94=B1ModelHub=20XC=E7=A4=BE=E5=8C=BA=E6=8F=90?= =?UTF-8?q?=E4=BE=9B=E6=A8=A1=E5=9E=8B?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Model: NightPrince/Muslim-6B-PRO Source: Original Platform --- .gitattributes | 38 +++++++ README.md | 189 +++++++++++++++++++++++++++++++++ chat_template.jinja | 54 ++++++++++ config.json | 89 ++++++++++++++++ generation_config.json | 7 ++ model.safetensors | 3 + muslim-6b-pro-banner-dark.png | 3 + muslim-6b-pro-banner-light.png | 3 + tokenizer.json | 3 + tokenizer_config.json | 15 +++ 10 files changed, 404 insertions(+) create mode 100644 .gitattributes create mode 100644 README.md create mode 100644 chat_template.jinja create mode 100644 config.json create mode 100644 generation_config.json create mode 100644 model.safetensors create mode 100644 muslim-6b-pro-banner-dark.png create mode 100644 muslim-6b-pro-banner-light.png create mode 100644 tokenizer.json create mode 100644 tokenizer_config.json diff --git a/.gitattributes b/.gitattributes new file mode 100644 index 0000000..a2d21f4 --- /dev/null +++ b/.gitattributes @@ -0,0 +1,38 @@ +*.7z filter=lfs diff=lfs merge=lfs -text +*.arrow filter=lfs diff=lfs merge=lfs -text +*.bin filter=lfs diff=lfs merge=lfs -text +*.bz2 filter=lfs diff=lfs merge=lfs -text +*.ckpt filter=lfs diff=lfs merge=lfs -text +*.ftz filter=lfs diff=lfs merge=lfs -text +*.gz filter=lfs diff=lfs merge=lfs -text +*.h5 filter=lfs diff=lfs merge=lfs -text +*.joblib filter=lfs diff=lfs merge=lfs -text +*.lfs.* filter=lfs diff=lfs merge=lfs -text +*.mlmodel filter=lfs diff=lfs merge=lfs -text +*.model filter=lfs diff=lfs merge=lfs -text +*.msgpack filter=lfs diff=lfs merge=lfs -text +*.npy filter=lfs diff=lfs merge=lfs -text +*.npz filter=lfs diff=lfs merge=lfs -text +*.onnx filter=lfs diff=lfs merge=lfs -text +*.ot filter=lfs diff=lfs merge=lfs -text +*.parquet filter=lfs diff=lfs merge=lfs -text +*.pb filter=lfs diff=lfs merge=lfs -text +*.pickle filter=lfs diff=lfs merge=lfs -text +*.pkl filter=lfs diff=lfs merge=lfs -text +*.pt filter=lfs diff=lfs merge=lfs -text +*.pth filter=lfs diff=lfs merge=lfs -text +*.rar filter=lfs diff=lfs merge=lfs -text +*.safetensors filter=lfs diff=lfs merge=lfs -text +saved_model/**/* filter=lfs diff=lfs merge=lfs -text +*.tar.* filter=lfs diff=lfs merge=lfs -text +*.tar filter=lfs diff=lfs merge=lfs -text +*.tflite filter=lfs diff=lfs merge=lfs -text +*.tgz filter=lfs diff=lfs merge=lfs -text +*.wasm filter=lfs diff=lfs merge=lfs -text +*.xz filter=lfs diff=lfs merge=lfs -text +*.zip filter=lfs diff=lfs merge=lfs -text +*.zst filter=lfs diff=lfs merge=lfs -text +*tfevents* filter=lfs diff=lfs merge=lfs -text +tokenizer.json filter=lfs diff=lfs merge=lfs -text +muslim-6b-pro-banner-dark.png filter=lfs diff=lfs merge=lfs -text +muslim-6b-pro-banner-light.png filter=lfs diff=lfs merge=lfs -text diff --git a/README.md b/README.md new file mode 100644 index 0000000..21f5a6a --- /dev/null +++ b/README.md @@ -0,0 +1,189 @@ +--- +license: apache-2.0 +language: + - ar + - en +base_model: Applied-Innovation-Center/Karnak-6B-v1.0 +pipeline_tag: text-generation +library_name: transformers +tags: + - text-generation + - causal-lm + - arabic + - islamic + - tool-calling + - lora + - peft + - qwen3 + - voice-assistant +--- + +

+ Muslim-6B-PRO +

+ +# Muslim-6B-PRO + +**Muslim-6B-PRO** is a behavior-tuned Islamic voice-assistant model, fine-tuned from +[Karnak-6B-v1.0](https://huggingface.co/Applied-Innovation-Center/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](https://huggingface.co/Applied-Innovation-Center/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": ""}, + {"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 `` 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](https://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](https://huggingface.co/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](https://huggingface.co/Applied-Innovation-Center/Karnak-6B-v1.0) +- Training dataset: [NightPrince/muslim-6b-v1-dataset](https://huggingface.co/datasets/NightPrince/muslim-6b-v1-dataset) +- GGUF quantizations: [NightPrince/Muslim-6B-PRO-GGUF](https://huggingface.co/NightPrince/Muslim-6B-PRO-GGUF) +- Live demo: [NightPrince/muslim-6b-pro-demo](https://huggingface.co/spaces/NightPrince/muslim-6b-pro-demo) +- Fine-tuning code: [github.com/NightPrinceY/Karnak-6B-Finetuning](https://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](https://www.apache.org/licenses/LICENSE-2.0) for the full text. Use of the +base model [Karnak-6B-v1.0](https://huggingface.co/Applied-Innovation-Center/Karnak-6B-v1.0) +remains subject to its own license terms from Applied Innovation Center. diff --git a/chat_template.jinja b/chat_template.jinja new file mode 100644 index 0000000..12b9069 --- /dev/null +++ b/chat_template.jinja @@ -0,0 +1,54 @@ +{%- if tools %} + {{- '<|im_start|>system\n' }} + {%- if messages[0]['role'] == 'system' %} + {{- messages[0]['content'] }} + {%- else %} + {{- 'You are Karnak, created by AIC. You are a helpful assistant.' }} + {%- endif %} + {{- "\n\n# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within XML tags:\n" }} + {%- for tool in tools %} + {{- "\n" }} + {{- tool | tojson }} + {%- endfor %} + {{- "\n\n\nFor each function call, return a json object with function name and arguments within XML tags:\n\n{\"name\": , \"arguments\": }\n<|im_end|>\n" }} +{%- else %} + {%- if messages[0]['role'] == 'system' %} + {{- '<|im_start|>system\n' + messages[0]['content'] + '<|im_end|>\n' }} + {%- else %} + {{- '<|im_start|>system\nYou are Qwen, created by Alibaba Cloud. You are a helpful assistant.<|im_end|>\n' }} + {%- endif %} +{%- endif %} +{%- for message in messages %} + {%- if (message.role == "user") or (message.role == "system" and not loop.first) or (message.role == "assistant" and not message.tool_calls) %} + {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>' + '\n' }} + {%- elif message.role == "assistant" %} + {{- '<|im_start|>' + message.role }} + {%- if message.content %} + {{- '\n' + message.content }} + {%- endif %} + {%- for tool_call in message.tool_calls %} + {%- if tool_call.function is defined %} + {%- set tool_call = tool_call.function %} + {%- endif %} + {{- '\n\n{"name": "' }} + {{- tool_call.name }} + {{- '", "arguments": ' }} + {{- tool_call.arguments }} + {{- '}\n' }} + {%- endfor %} + {{- '<|im_end|>\n' }} + {%- elif message.role == "tool" %} + {%- if (loop.index0 == 0) or (messages[loop.index0 - 1].role != "tool") %} + {{- '<|im_start|>user' }} + {%- endif %} + {{- '\n\n' }} + {{- message.content }} + {{- '\n' }} + {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %} + {{- '<|im_end|>\n' }} + {%- endif %} + {%- endif %} +{%- endfor %} +{%- if add_generation_prompt %} + {{- '<|im_start|>assistant\n' }} +{%- endif %} diff --git a/config.json b/config.json new file mode 100644 index 0000000..838325f --- /dev/null +++ b/config.json @@ -0,0 +1,89 @@ +{ + "architectures": [ + "Qwen3ForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.0, + "bos_token_id": null, + "dtype": "float16", + "eos_token_id": 151645, + "head_dim": 128, + "hidden_act": "silu", + "hidden_size": 2560, + "initializer_range": 0.02, + "intermediate_size": 9728, + "layer_types": [ + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention" + ], + "max_position_embeddings": 262144, + "max_window_layers": 36, + "model_type": "qwen3", + "num_attention_heads": 32, + "num_hidden_layers": 54, + "num_key_value_heads": 8, + "pad_token_id": 151643, + "rms_norm_eps": 1e-06, + "rope_parameters": { + "rope_theta": 5000000, + "rope_type": "default" + }, + "sliding_window": null, + "tie_word_embeddings": true, + "transformers_version": "5.12.1", + "use_cache": false, + "use_sliding_window": false, + "vocab_size": 192728 +} diff --git a/generation_config.json b/generation_config.json new file mode 100644 index 0000000..0350fa4 --- /dev/null +++ b/generation_config.json @@ -0,0 +1,7 @@ +{ + "_from_model_config": true, + "eos_token_id": 151645, + "pad_token_id": 151643, + "transformers_version": "5.12.1", + "use_cache": false +} diff --git a/model.safetensors b/model.safetensors new file mode 100644 index 0000000..5e90729 --- /dev/null +++ b/model.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:410f75d5e7eaec715ee94db04631b71a1bbabe8f70a8be854a8034c5b07af636 +size 11887369032 diff --git a/muslim-6b-pro-banner-dark.png b/muslim-6b-pro-banner-dark.png new file mode 100644 index 0000000..58641c6 --- /dev/null +++ b/muslim-6b-pro-banner-dark.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e712e4e0fd118755ed240ffaa4a1f52b308e8003cb95478fdad9c5731bc98daf +size 238301 diff --git a/muslim-6b-pro-banner-light.png b/muslim-6b-pro-banner-light.png new file mode 100644 index 0000000..849cf69 --- /dev/null +++ b/muslim-6b-pro-banner-light.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b4478010053dc0bd9af322f60c44f6b1d0da1181c800f6e2324527501f58ff19 +size 232878 diff --git a/tokenizer.json b/tokenizer.json new file mode 100644 index 0000000..57cbdae --- /dev/null +++ b/tokenizer.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f644d71501367aaa569353e4787cd25cadb851169b709d8712bff51c72ce76f8 +size 15658511 diff --git a/tokenizer_config.json b/tokenizer_config.json new file mode 100644 index 0000000..9158804 --- /dev/null +++ b/tokenizer_config.json @@ -0,0 +1,15 @@ +{ + "add_prefix_space": false, + "backend": "tokenizers", + "bos_token": null, + "clean_up_tokenization_spaces": false, + "eos_token": "<|im_end|>", + "errors": "replace", + "is_local": true, + "local_files_only": false, + "model_max_length": 131072, + "pad_token": "<|endoftext|>", + "split_special_tokens": false, + "tokenizer_class": "Qwen2Tokenizer", + "unk_token": null +}