Upload README.md with huggingface_hub

README.md

@@ -3,12 +3,12 @@ license: other
 license_name: hayula-research-license-v1
 license_link: https://hayula.xyz/license
 language:
-- en
 - ar
+- en
 tags:
-- bug-bounty
-- security
-- saif
+- arabic
+- bilingual
+- llama
 - lora
 - qwen2.5
 - averroes
@@ -17,34 +17,85 @@ pipeline_tag: text-generation
 base_model: hayulalab/Averroes-Q-Instruct
 ---
 
-# Hayula-Kindi-7B-LoRA
+<p align="center">
+  <img src="https://huggingface.co/hayulalab/assets/resolve/main/banner.jpg" alt="Hayula AI Lab" width="100%">
+</p>
+
+# كندي / Hayula-Kindi-7B-LoRA
+
+**اختبار الاختراق** — 309 ثنائي تحليل أمني، 55 هدف
+
+---
+
+## English
+
+### Hayula-Kindi-7B-LoRA
+
+**Bug Bounty Specialist** — LoRA adapter fine-tuned on SAIF security analysis data, 309 pairs, 55 targets.
+
+Named after Al-Kindi, the father of Arab cryptography and code-breaking.
+
+| Property | Value |
+|:---------|:------|
+| **Base Model** | Averroes-Q-Instruct (Qwen2.5-7B) |
+| **Method** | LoRA (rank 8, scale 20, 16 layers) |
+| **Training** | 500 iterations, lr 1e-5, batch 4 |
+| **Hardware** | Apple M2 Ultra (192GB) |
+| **Adapter Size** | 44MB |
+| **Metrics** | Test loss 0.091, PPL 1.095 |
+
+### Quick Start
+
+```python
+from mlx_lm import load, generate
+from mlx_lm.lora import load_adapters
 
-Bug Bounty Specialist — LoRA adapter trained on SAIF security analysis data.
+model, tokenizer = load("hayulalab/Averroes-Q-Instruct")
+load_adapters(model, "hayulalab/Hayula-Kindi-7B-LoRA")
 
-Named after Al-Kindi, father of Arab cryptography.
+messages = [{"role": "user", "content": "Analyze this finding..."}]
+prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True)
+response = generate(model, tokenizer, prompt=prompt, max_tokens=256)
+print(response)
+```
 
-## Model Details
+---
 
-- **Base Model:** Averroes-Q-Instruct (Qwen2.5-7B)
-- **Method:** LoRA (rank 8, scale 20, 16 layers)
-- **Data:** 309 SAIF analysis pairs (55 targets)
-- **Adapter:** 44MB
-- **Test Loss:** 0.091 | PPL: 1.095
+## العربية
 
-## Targets
+### كندي — نموذج اختبار الاختراق
 
-Apple, Coinbase, Coupang, DoorDash, Meta, GitLab, Google, MongoDB,
-NetScaler, Shopify, Snapchat, Uber, Adobe, Airbnb, Atlassian, Wickr
+مُدرَّب باستخدام LoRA على بيانات تحليل الثغرات الأمنية من منصة SAIF.
 
-## Usage
+سُمّي على اسم أبو يوسف يعقوب الكندي، رائد التشفير العربي.
+
+| الخاصية | القيمة |
+|:--------|:-------|
+| **النموذج الأساسي** | Averroes-Q-Instruct |
+| **طريقة التدريب** | LoRA (rank 8, scale 20, 16 layers) |
+| **عدد التكرارات** | 500 |
+| **معدل التعلم** | 1e-5 |
+| **الجهاز** | Apple M2 Ultra (192GB) |
+| **حجم المحوّل** | 44MB |
+| **النتائج** | Test loss 0.091, PPL 1.095 |
+
+### البدء السريع
 
 ```python
 from mlx_lm import load, generate
 from mlx_lm.lora import load_adapters
+
 model, tokenizer = load("hayulalab/Averroes-Q-Instruct")
 load_adapters(model, "hayulalab/Hayula-Kindi-7B-LoRA")
+
+messages = [{"role": "user", "content": "حلل هذه الثغرة الأمنية..."}]
+prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True)
+response = generate(model, tokenizer, prompt=prompt, max_tokens=256)
+print(response)
 ```
 
-## License
+---
+
+### License / الترخيص
 
-Hayula Research License v1.0
+Hayula Research License v1.0 — [Full terms](https://hayula.xyz/license)