初始化项目，由ModelHub XC社区提供模型
Model: NightPrince/Muslim-6B-PRO Source: Original Platform
This commit is contained in:
ModelHub XC
38
.gitattributes
vendored
Normal file
38
.gitattributes
vendored
Normal file
@@ -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
|
||||
189
README.md
Normal file
189
README.md
Normal file
@@ -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
|
||||
---
|
||||
|
||||
<p align="center">
|
||||
<img src="https://huggingface.co/NightPrince/Muslim-6B-PRO/resolve/main/muslim-6b-pro-banner-light.png" alt="Muslim-6B-PRO" width="100%" />
|
||||
</p>
|
||||
|
||||
# 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": "<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](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.
|
||||
54
chat_template.jinja
Normal file
54
chat_template.jinja
Normal file
@@ -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 <tools></tools> XML tags:\n<tools>" }}
|
||||
{%- for tool in tools %}
|
||||
{{- "\n" }}
|
||||
{{- tool | tojson }}
|
||||
{%- endfor %}
|
||||
{{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|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<tool_call>\n{"name": "' }}
|
||||
{{- tool_call.name }}
|
||||
{{- '", "arguments": ' }}
|
||||
{{- tool_call.arguments }}
|
||||
{{- '}\n</tool_call>' }}
|
||||
{%- endfor %}
|
||||
{{- '<|im_end|>\n' }}
|
||||
{%- elif message.role == "tool" %}
|
||||
{%- if (loop.index0 == 0) or (messages[loop.index0 - 1].role != "tool") %}
|
||||
{{- '<|im_start|>user' }}
|
||||
{%- endif %}
|
||||
{{- '\n<tool_response>\n' }}
|
||||
{{- message.content }}
|
||||
{{- '\n</tool_response>' }}
|
||||
{%- 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 %}
|
||||
89
config.json
Normal file
89
config.json
Normal file
@@ -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
|
||||
}
|
||||
7
generation_config.json
Normal file
7
generation_config.json
Normal file
@@ -0,0 +1,7 @@
|
||||
{
|
||||
"_from_model_config": true,
|
||||
"eos_token_id": 151645,
|
||||
"pad_token_id": 151643,
|
||||
"transformers_version": "5.12.1",
|
||||
"use_cache": false
|
||||
}
|
||||
3
model.safetensors
Normal file
3
model.safetensors
Normal file
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:410f75d5e7eaec715ee94db04631b71a1bbabe8f70a8be854a8034c5b07af636
|
||||
size 11887369032
|
||||
3
muslim-6b-pro-banner-dark.png
Normal file
3
muslim-6b-pro-banner-dark.png
Normal file
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:e712e4e0fd118755ed240ffaa4a1f52b308e8003cb95478fdad9c5731bc98daf
|
||||
size 238301
|
||||
3
muslim-6b-pro-banner-light.png
Normal file
3
muslim-6b-pro-banner-light.png
Normal file
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:b4478010053dc0bd9af322f60c44f6b1d0da1181c800f6e2324527501f58ff19
|
||||
size 232878
|
||||
3
tokenizer.json
Normal file
3
tokenizer.json
Normal file
@@ -0,0 +1,3 @@
|
||||
version https://git-lfs.github.com/spec/v1
|
||||
oid sha256:f644d71501367aaa569353e4787cd25cadb851169b709d8712bff51c72ce76f8
|
||||
size 15658511
|
||||
15
tokenizer_config.json
Normal file
15
tokenizer_config.json
Normal file
@@ -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
|
||||
}
|
||||
Reference in New Issue
Block a user