Supernova11c/Supernova-Roman-teraillm-Restoration-V3.1

Supernova Roman Restoration V3.1

A deterministic, rule-based Romanized Nepali → Devanagari restoration system developed by Supernova.

V3.1 is designed specifically for Romanized Nepali. It does not use a transformer, pretrained model, neural network, embeddings, or generative language model.

The system relies on large-scale linguistic evidence, deterministic dictionary lookup, contextual rules, and conservative fallback behavior.

---

🚀 Highlights

• 100% exact accuracy on the reported 198,558-example held-out test set
• 0 incorrect predictions on that test set
• 1.58M+ training examples
• 2.17M+ Roman→native mappings available from the V2 system
• Context-aware restoration rules
• Deterministic output
• CPU-friendly
• No neural inference
• No pretrained parameters
• No transformer
• No external generative model

---

Architecture

                 Romanized Nepali Input
                           │
                           ▼
                 ┌─────────────────────┐
                 │ Exact Phrase Lookup │
                 └──────────┬──────────┘
                            │
                            ▼
                 ┌─────────────────────┐
                 │ Contextual Rule     │
                 │ Engine              │
                 └──────────┬──────────┘
                            │
                            ▼
                 ┌─────────────────────┐
                 │ Training-Derived    │
                 │ Word Evidence       │
                 └──────────┬──────────┘
                            │
                            ▼
                 ┌─────────────────────┐
                 │ V2 Word / Dictionary│
                 │ Lookup              │
                 └──────────┬──────────┘
                            │
                            ▼
                 ┌─────────────────────┐
                 │ Conservative       │
                 │ Fallback            │
                 └──────────┬──────────┘
                            │
                            ▼
                     Devanagari Output