km-bpe-md-v3-spm

🇰🇭 Khmer BPE-MD-v3-SPM (Hybrid Tokenizer)

BPE-MD-v3-SPM is a hybrid Khmer tokenizer that combines:

• BPE-MD (Morphology-Driven) rules for Khmer word segmentation, and
• SentencePiece BPE modeling for subword learning, coverage, and byte safety.

This tokenizer is built for both Khmer and bilingual (Khmer + English + Math) text. It handles Unicode normalization, symbols, and numerics gracefully — ideal for LLMs, translation models, or RAG systems.

---

🧠 Features

• Hybrid design: BPE-MD (morphology) × SentencePiece (subword)
• Script coverage: Khmer + Latin + Math + Digits
• Vocab size: 16 100
• Character coverage: 1.0
• Includes: user-defined math and chemical tokens (√, ², ₁₀, H₂O, log₁₀, etc.)

---

🧩 Example usage

(from transformers import T5Tokenizer)

(tok = T5Tokenizer.from_pretrained("Msok99/km-bpe-md-v3-spm"))

(text = "ខ្ញុំបានគណនាថា √25 + 3² = 34")

(print(tok.tokenize(text)))

(print(tok.decode(tok.encode(text))))

---

📊 Training details

• Base: Khmer Morphology-Driven corpus (education, news, QA)
• Algorithm: SentencePiece (BPE mode)
• User symbols: Mathematical, scientific, and Khmer-digit patterns
• Goal: Robust tokenization for LLM fine-tuning on Khmer + mixed-script data

---

📜 License

MIT