SentencePiece-BPE tokenizer (exact vocabulary size 64,000) designed for NyayaLM-200M, a bilingual (Nepali + English) legal-domain language model.

It is optimized for:

Native Nepali (Devanagari)
Romanized Nepali (Latin script)
English
Mixed code-switched legal / everyday text
ChatML-style conversation formatting (Qwen3)

Key Features

Feature	Value
Vocabulary size	64,000
Algorithm	SentencePiece BPE + byte fallback
Character coverage	0.9995
Special tokens	ChatML (`<
Pre-tokenizer	Metaspace (`▁`)
Normalization	Identity (no aggressive NFKC)
Split digits / numbers	Yes
Split by Unicode script	Yes
Byte fallback	Enabled

The tokenizer was trained without any hardcoded seed symbols — the entire alphabet (Devanagari + Latin + digits + punctuation) is learned purely from the bilingual legal + general corpus.

Training Data Summary

A carefully balanced bilingual corpus was constructed from the following sources (streaming + deduplication via BLAKE2b):

Nepali (Devanagari)

Wikipedia Nepali dump
Himalaya-AI Nepali tokenizer corpus
Nepali Supreme Court judgement text corpus
Sangraha Nepali
Nepal Legal Acts corpus (Nepali)

≈ 885k documents · ~1.33 billion characters

English + Romanized Nepali

GPT-2 style English pre-train corpus
High-quality English sentences
Nepal Legal Acts corpus (English)
WikiText-103 (raw)
Roman-Nepali Alpaca
Roman-Nepali e-commerce QA (ShareGPT format)

≈ 2.40 million documents · ~1.22 billion characters

Final training corpus after balancing & interleaving:

3.28 million lines
~3.9 GB of clean text
Strong coverage of both Devanagari and Latin scripts (verified by character statistics)

Intended Use

Pre-training / continued pre-training of small–medium LLMs focused on Nepali legal text
Instruction tuning and chat with ChatML format
Any application that needs efficient tokenization of Nepali + English + Romanized Nepali mixtures

Evaluation Metrics

We evaluated the tokenizer against strong baselines on four language categories:

Pure English
Pure Nepali (Devanagari)
Romanized Nepali
Mixed / code-switched

Primary metrics (lower fertility & higher STRR are better):

Metric	Description	Direction
Fertility	Tokens per whitespace-delimited word	↓ better
Characters per token	Compression efficiency	↑ better
STRR	Single-Token Retention Rate (% of words kept intact)	↑ better

Overall Mean Metrics (all languages combined)

Tokenizer	Fertility ↓	Chars/Token ↑	STRR ↑
Custom (nyayalm-64k)	1.483	4.294	0.709
Gemma-4	1.536	4.102	0.495
GPT-OSS-20B (o200k_harmony)	1.600	3.962	0.448
GPT-5 (o200k)	1.600	3.962	0.448
Llama-3.2	2.102	3.265	0.414
Qwen3.5	2.329	3.126	0.408
Claude (approx)	3.252	2.779	0.393

Per-Language Fertility (lower = better)

Tokenizer	English	Mixed	Nepali	Roman_Nepali
Custom (nyayalm-64k)	1.131	1.135	1.252	2.417
Gemma-4	1.131	1.099	1.769	2.147
GPT-5 (o200k)	1.147	1.157	1.823	2.272
GPT-OSS-20B (o200k_harmony)	1.147	1.157	1.823	2.272
Llama-3.2	1.189	1.471	3.311	2.437
Qwen3.5	1.189	1.587	4.179	2.362
Claude (approx)	1.172	1.984	7.277	2.573

Per-Language STRR (higher = better)

Tokenizer	English	Mixed	Nepali	Roman_Nepali
Custom (nyayalm-64k)	0.869	0.865	0.816	0.283
Gemma-4	0.756	0.768	0.243	0.213
GPT-5 (o200k)	0.675	0.712	0.218	0.188
GPT-OSS-20B (o200k_harmony)	0.675	0.712	0.218	0.188
Llama-3.2	0.675	0.708	0.084	0.188
Qwen3.5	0.675	0.708	0.059	0.188
Claude (approx)	0.656	0.688	0.040	0.188

Key takeaways:

Best overall fertility (1.483) and highest STRR (0.709) among all compared tokenizers.
Dramatically lower fertility on pure Nepali (1.252 vs 1.77–7.28 for others).
Highest single-token retention on Nepali (81.6%) — most words stay intact.
Competitive or best on English and Mixed text.
Strong advantage on legal-domain and code-switched Nepali–English sentences.

Usage

python

from transformers import AutoTokenizer

tok = AutoTokenizer.from_pretrained("chhatramani/nyayalm-tokenizer-64k_v2")

# Simple encode / decode
text = "नेपालको संविधान २०७२ ले नागरिकलाई मौलिक हक प्रदान गरेको छ।"
ids = tok.encode(text, add_special_tokens=False)
print(len(ids), tok.decode(ids))

# ChatML conversation
messages = [
    {"role": "system", "content": "You are a helpful Nepali legal assistant."},
    {"role": "user", "content": "मेरो मौलिक हक के के हुन्?"},
]
prompt = tok.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
print(prompt)

Special token IDs (example)

python

Pythonprint(tok.convert_tokens_to_ids("<|endoftext|>"))  # 63982
print(tok.convert_tokens_to_ids("<|im_start|>"))   # 63983
print(tok.convert_tokens_to_ids("<|im_end|>"))     # 63984
print(len(tok))

Citation

If you use this tokenizer in your research or product, please cite:

bibtex

bibtex@misc{nyayalm-tokenizer-64k-v2,
  title  = {NyayaLM 64k SentencePiece-BPE Tokenizer},
  author = {Chhatramani},
  year   = {2026},
  url    = {https://huggingface.co/chhatramani/nyayalm-tokenizer-64k_v2}
}

License

Apache License 2.0