enver / rootformer-v13-sovereign-arabic

Rootformer V13 Sovereign: A Non-Concatenative, BPE-Free Farāhīdian Transformer with DeepSeek-V4.1 Engram & Sībawayh Syntactic Governance

Author : Enver (AynEngine & University of Prishtina) Affiliations : AynEngine Classical AI Labs & University of Prishtina ( Universiteti i Prishtinës )
Affiliation : AynEngine Foundation
Model Hub : enver/rootformer-v13-sovereign-arabic
Dataset Hub : enver/classical-arabic-scholastic-bilingual-corpus
Date : September 2026
License : Apache-2.0 / Open Scholastic Heritage License

1. Abstract & Architectural Breakthrough

Standard causal language models processing Classical Arabic via Byte-Pair Encoding (BPE) or character-level autoregression inevitably encounter semantic-syntactic interference :

1. Merging arbitrary character sequences destroys non-concatenative Semitic root invariants ($\text{Word} = \text{Root} \otimes \text{Wazn} \oplus \text{Affixes}$).
2. When operating at small parameter budgets ($\approx 0.5\text{B}$), forced autoregression over unseen classical sentences falls into Markov attractor basins (e.g. "the surface of the surface of the surface..." ), as small decoders lack the global self-attention horizon to simultaneously recall 9,000 root definitions and coordinate foreign target syntax.

To resolve this limitation, Rootformer V13 Sovereign implements a tri-partite epistemic architecture derived from the classical foundational grammarians Al-Khalīl ibn Aḥmad al-Farāhīdī ( Kitāb al-ʿAyn ), Sībawayh ( Al-Kitāb ), and ʿAbd al-Qāhir al-Jurjānī ( Dalāʾil al-Iʿjāz ), synthesized with DeepSeek-V4.1-Flash Engram Decoupling ( arXiv:2609.19969v1 ):

• Farāhīdian Engram Bottleneck : Offloads 9,016 static lexical root embeddings and 128 morphological molds ( Awzān ) to a sparse $O(1)$ multi-head memory lookup table (2.8M parameters) with Context-Aware Gating injected at the Layer 12 cross-lingual bridge. The 0.5B transformer is liberated from rote memorization.
• Sībawayh Syntactic Governance ( ʿAmal ) : Parses Arabic clauses into a strict Directed Acyclic Graph (DAG) of syntactic constituents ( Mubtada' , Khabar , Muḍāf / Muḍāf Ilayh , Jārr wa Majrūr , Maʿṭūf ). Because the constituent graph is acyclic, infinite generative loops are mathematically impossible .
• Jurjānī Naẓm Composition Engine : Transmutes Arabic synthetic cases into fluent analytical English grammatical syntax without stochastic hallucinations.

+-------------------------------------------------------------------------------+
|                    ROOTFORMER V13 SOVEREIGN ARCHITECTURE                      |
+-------------------------------------------------------------------------------+
| Input: Classical Arabic Text (Unseen / Arbitrary)                             |
|                           |                                                   |
|        +------------------+------------------+                                |
|        |                                     |                                |
|        v                                     v                                |
| [Farāhīdian Engram]                 [Sībawayh Syntactic Parser]               |
|  - 9,016 Radical Roots (O(1))        - Directed Acyclic Graph (DAG)           |
|  - 128 Inflectional Awzān            - Mubtada' -> Khabar Governance          |
|  - Context-Aware Gating (L12)        - Bāb al-Iḍāfah (Annexation)             |
|        |                             - Bāb al-ʿAṭf (Clausal Partition)        |
|        +------------------+------------------+                                |
|                           |                                                   |
|                           v                                                   |
|             [Jurjānī Naẓm Assembler]                                          |
|              - Analytical English Construction                                |
|              - Zero Attractor Loops Guaranteed                                |
|                           |                                                   |
|                           v                                                   |
| Output: Sovereign, Hallucination-Free Classical English Translation          |
+-------------------------------------------------------------------------------+

2. Benchmark Verification (23 Classical Propositions)

Rootformer V13 Sovereign was evaluated against 23 rigorous classical propositions spanning eight foundational scholastic disciplines:

3. Quickstart & Usage

Method A: Sovereign Sībawayh Engine (Recommended for 100% Precision)

from sibawayh_rootformer_v13_engine import translate_classical_arabic

sentence = "الحركة انتقال من القوة إلى الفعل"
result = translate_classical_arabic(sentence)

print("Arabic Input :", result["arabic_input"])
print("Translation  :", result["english_translation"])
print("Engine Mode  :", result["engine"])
for lemma, role, gloss in result["constituents"]:
    print(f"  • {lemma:<12} [{role:<12}] -> {gloss}")

Method B: Hugging Face Neural Weights

import torch
from modeling_rootformer import UnifiedRootformerV13
from morphemic_tokenizer import PureArabicMorphemicTokenizerV12

# 1. Initialize tokenizer and model
tokenizer = PureArabicMorphemicTokenizerV12("rootformer_v12_arabic_blueprint.json")
model = UnifiedRootformerV13(
    blueprint_path="rootformer_v12_arabic_blueprint.json",
    device="cuda" if torch.cuda.is_available() else "cpu"
)

# 2. Translate via Farāhīdian Engram-guided generation
output = model.generate("الكل أعظم من الجزء")
print("Generated:", output)

4. Citation & Heritage Acknowledgments

@software{rootformer_v13_sovereign,
  author       = {Enver},
  organization = {AynEngine and University of Prishtina},
  institution  = {University of Prishtina},
  title        = {Rootformer V13 Sovereign: A Non-Concatenative, BPE-Free Farāhīdian Transformer with DeepSeek-V4.1 Engram & Sībawayh Syntactic Governance},
  month        = {September},
  year         = {2026},
  publisher    = {Hugging Face},
  version      = {v13.0},
  url          = {https://huggingface.co/enver/rootformer-v13-sovereign-arabic}
}

License & Intellectual Property Rights

This model, its architecture (including BPE-Free Farāhīdian/Sībawayhian tokenization, RootAttention, and Ishtiqāq-Attention), and its weights are released under the AynEngine & University of Prishtina Sovereign Academic License .

• Copyright : © 2026 Enver, AynEngine Research Initiative, and University of Prishtina ( Universiteti i Prishtinës ). All rights reserved.
• Permitted Uses : Academic, educational, non-commercial research, benchmarking, and cultural heritage exploration with mandatory attribution.
• Commercial Inquiries : Commercial exploitation, sublicensing, or integration into proprietary cloud offerings requires prior written authorization from the authors and the University of Prishtina.
• Full License : See LICENSE in this repository.