enver / ayncoding-qwen2.5-coder-1.5b

🏛️ AynCoding: Sovereign Epistemic Classical Arabic Logic (Manṭiq) & Morphology 1.5B

AynCoding ( ayncoding-qwen2.5-coder-1.5b ) is a sovereign, epistemic code synthesis and architectural refactoring model based on Qwen2.5-Coder-1.5B , conditioned and aligned with Classical Arabic Logic ( Manṭiq ) and Morphological Root Lexicography ( Ishtiqāq ) .

Unlike standard LLMs trained on unstructured, messy GitHub repositories, AynCoding natively engages an epistemic Chain-of-Thought reasoning block ( <ayn_mantiq> ... </ayn_mantiq> ) prior to code generation. It analyzes software requirements through ontological definitions ( Al-Ḥadd ), eliminates logical fallacies ( Dafʿ al-Dawr , Dafʿ al-Tasalsul , ʿAdam al-Tanāquḍ ), and maps software engineering domains to classical tri-consonantal roots.

🌟 Governing Classical Authorities

AynCoding is grounded in seven classical Arabic masterpieces:

1. Abū Ḥāmid al-Ghazālī ( Miʿyār al-ʿIlm fī Fann al-Manṭiq & Miḥakk al-Naẓar ):
   • Al-Ḥadd bi al-Dhātiyyāt (Real Definition by Essence) : Every class, struct, and routine must be defined by its essential constitutive attributes, not superficial symptoms.
   • Dafʿ al-Dawr (Elimination of Circularity) : Complete prohibition of circular dependencies, circular imports, and mutual deadlocks.
   • Dafʿ al-Tasalsul (Elimination of Infinite Regress) : Strictly bounded recursion, guaranteed loop termination, and timeout bounds.
   • ʿAdam al-Tanāquḍ (Law of Non-Contradiction) : Mutually exclusive state FSMs; elimination of contradictory states and silent exception swallowing.
2. Fakhr al-Dīn al-Rāzī ( Al-Mulakhkhaṣ fī al-Ḥikmah wa al-Manṭiq ): Proof theory, syllogistic validation, and formal deductive rigor.
3. Al-Khalīl ibn Aḥmad al-Farāhīdī ( Kitāb al-ʿAyn ):
   • Atomic Primitive Decomposition : Decomposing complex software architectures into orthogonal, irreducible primitives.
4. Ibn Manẓūr ( Lisān al-ʿArab ):
   • Exhaustive Error Taxonomy & State-Space Coverage : Zero unhandled match arms, zero silent except Exception: pass , and complete lifecycle modeling ( INITIALIZING -> ACTIVE -> DEGRADED -> CLOSED -> FAILED ).
5. Al-Rāghib al-Iṣfahānī ( Al-Mufradāt fī Gharīb al-Qurʾān ):
   • Ontological Domain Modeling & Teleology ( Ghāyah ) : Absolute ban on amorphous, vague identifiers ( data , temp , val , mgr , helper ). Every token represents its distinct purpose.
6. Al-Zamakhsharī ( Asās al-Balāghah ):
   • Abstraction Integrity & Rhetorical Eloquence ( Ḥaqīqah vs Majāz ) : Distinguishing physical hardware reality (memory, CPU caches, IO) from metaphors. Zero leaky abstractions.
7. Sībawayh ( Al-Kitāb ):
   • Syntactic Governance ( Al-ʿĀmil wa al-Maʿmūl ) : Strict caller-callee hierarchy, rigorous static typing, and guaranteed AST integrity.

🧠 Cognitive Chain-of-Thought Format ( <ayn_mantiq> )

Before producing code, AynCoding generates an epistemic reasoning block:

<ayn_mantiq>
🏛️ AYN-ENGINE EPISTEMIC LOGIC & MORPHOLOGY REASONING:
- Classical Root & Morphology (الجذر والتصريف): [e.g. ق-ف-ل for locking / mutual exclusion]
- Real Definition & Essence (الحد بالذاتيات - معيار العلم للغزالي): [Essential attributes and invariants]
- Epistemic Fallacy Invariants (دفع الدور والتسلسل ونفي التناقض): [Circularity, regress, and contradiction guards]
- Lexicographical Teleology (الغاية وبلاغة التجريد - المفردات والأساس): [Pure purpose, zero vague abstractions]
- Syntactic Governance (العامل والمعمول - كتاب سيبويه): [Strict type contracts, caller-callee hierarchy]
</ayn_mantiq>

📊 Head-to-Head Epistemic Benchmark vs. Frontier AI

Evaluated blindly and deterministically using the 5-Pillar Static Epistemic Auditor across 3 canonical systems challenges:

AynCoding 1.5B running on pure CPU trails the frontier cloud AI by only 2.5% while achieving a perfect 100% on finite-state governance.

🚀 Quickstart: Running with Ollama

1. Download the GGUF model and Modelfile:

   huggingface-cli download enver/ayncoding-qwen2.5-coder-1.5b ayncoding-qwen2.5-coder-1.5b.gguf Modelfile --local-dir ./ayncoding
   cd ayncoding
   

2. Register with Ollama:

   ollama create ayncoding-model -f Modelfile
   

3. Run Inference:

   ollama run ayncoding-model "Write a thread-safe token bucket rate limiter in Python"
   

💻 Python / Transformers & llama.cpp Usage

With llama-cpp-python:

from llama_cpp import Llama

llm = Llama(
    model_path="ayncoding-qwen2.5-coder-1.5b.gguf",
    n_ctx=8192,
    n_threads=16  # Pin to single socket for dual-Xeon CPU optimization
)

output = llm.create_chat_completion(
    messages=[
        {"role": "user", "content": "Implement an in-memory ring buffer with atomic sequence counters in Python."}
    ],
    temperature=0.2,
    max_tokens=4096
)

print(output["choices"][0]["message"]["content"])

🛠️ Included Tools in Repository

• Modelfile : Sovereign Ollama definition with the 5 classical Arabic pillars and epistemic system prompt.
• dataset/ayn_mantiq_epistemic_dataset_50.jsonl : 50 AST-validated training samples across 10 classical software engineering domains with explicit <ayn_mantiq> CoT reasoning.
• adapter/adapter_manifest.json : CPU LoRA fine-tuning configuration manifest (Rank 16, Alpha 32).
• training/train_mantiq_lora.py : PyTorch PEFT CPU fine-tuning script with NUMA thread optimization.
• tools/core/mantiq_purifier.py : Automated code sanitizer purging amorphous variables ( data , temp , val , mgr ) and silent bare except: clauses.
• tools/core/mantiq_engine.py : Classical Arabic Logic fallacy engine detecting Dawr , Tasalsul , and Tanāquḍ .

📜 Citation & Heritage

If you use this model or dataset, please cite the classical authorities:

• Abū Ḥāmid al-Ghazālī, Miʿyār al-ʿIlm fī Fann al-Manṭiq (Criterion of Knowledge in Logic).
• Al-Khalīl ibn Aḥmad al-Farāhīdī, Kitāb al-ʿAyn (The First Arabic Lexicon).
• Al-Rāghib al-Iṣfahānī, Al-Mufradāt fī Gharīb al-Qurʾān (Ontological Lexicon).
• Sībawayh, Al-Kitāb (The Foundation of Arabic Grammar & Governance).