---
license: cc-by-nc-nd-4.0
language:
- en
- hi
- kn
- ta
- te
- ml
- mr
- es
- zh
- ar
- ru
pipeline_tag: token-classification
tags:
- tokenizer
- bpe
- multilingual
- indic-languages
- luna
- text-generation
---

# LUNA 1B Tokenizer (Asterizer)

The **multilingual BPE tokenizer family** for the LUNA-1B model — a custom, Indic-aware tokenizer ("Asterizer") trained on the 240 GB LUNA multilingual corpus. It ships in three vocabulary sizes: **128K**, **64K**, and **8K**.

This is the successor to the original [LUNA-Tokenizer](https://huggingface.co/ASTERIZER/LUNA-Tokenizer) (EleutherAI/Pythia-160m GPT-NeoX, vocab 50,304), which powers the smaller LUNA-100M / LUNA-300M models.

## Why a new tokenizer

The original GPT-NeoX tokenizer — like `cl100k_base` — is **English-optimized**. On Indic scripts (Kannada, Tamil, Telugu, Malayalam, Hindi) it fragments a single script character into multiple tokens, bloating sequence length and wasting compute. The Asterizer tokenizer is trained on a multilingual corpus, so it compresses Indic text dramatically better **while staying at parity on English and code**.

---

## Benchmark — multilingual compression

Measured on a fixed **15-language probe set** (English, Hindi, Kannada, Tamil, Telugu, Malayalam, Marathi, Kanglish, Python, JavaScript, Spanish, Chinese, Arabic, Russian) — **2,144 bytes total**. **Lower tokens = better.**

### Totals

| Tokenizer | Vocab | Total tokens | Bytes/token | vs GPT-NeoX |
|---|---|---:|---:|---:|
| **GPT-NeoX 50K (old LUNA)** | 50,304 | 945 | 2.27 | 1.00× (baseline) |
| **Asterizer 128K** | 128K | **335** | **6.40** | **2.82× (−65%)** |
| **Asterizer 64K** | 64K | **392** | **5.47** | **2.41× (−58%)** |
| **o200k_base** (GPT-4o) | 200K | 350 | 6.13 | 2.70× (−63%) |
| **cl100k_base** (GPT-3.5/4) | 100K | 925 | 2.32 | 1.02× (−2%) |

**Headline:** Asterizer 128K compresses multilingual text **2.8× better** than the old LUNA tokenizer, and even **edges out OpenAI's o200k_base** (335 vs 350 tokens) on this Indic-heavy set. `cl100k_base` is just as poor as the old tokenizer on Indic scripts.

### Per-language tokens (lower = better)

| Language | GPT-NeoX 50K | Asterizer 128K | Asterizer 64K | o200k_base | cl100k_base |
|---|---:|---:|---:|---:|---:|
| English (general) | 18 | 20 | 22 | 17 | 18 |
| English (technical) | 19 | 19 | 22 | 18 | 18 |
| Hindi | 79 | **25** | 30 | 23 | 81 |
| **Kannada** | 185 | **26** | 34 | 37 | 185 |
| **Tamil** | 107 | **20** | 22 | 30 | 108 |
| **Telugu** | 109 | **18** | 21 | 31 | 122 |
| **Malayalam** | 134 | **20** | 23 | 29 | 135 |
| Marathi | 69 | 26 | 32 | 26 | 67 |
| Kanglish (Latin-script) | 29 | 27 | 27 | 25 | 26 |
| Python | 33 | 29 | 31 | 26 | 25 |
| JavaScript | 27 | 27 | 29 | 23 | 21 |
| Spanish | 22 | 19 | 24 | 15 | 19 |
| Chinese | 37 | 18 | 24 | 20 | 29 |
| Arabic | 39 | 20 | 22 | 15 | 39 |
| Russian | 38 | 21 | 29 | 15 | 32 |

### Indic-language improvement (old → 128K)

| Language | Old tokens | 128K tokens | Compression |
|---|---:|---:|---:|
| **Kannada** | 185 | 26 | **7.1× fewer** |
| **Malayalam** | 134 | 20 | **6.7× fewer** |
| **Telugu** | 109 | 18 | **6.1× fewer** |
| **Tamil** | 107 | 20 | **5.4× fewer** |
| **Hindi** | 79 | 25 | **3.2× fewer** |
| **Marathi** | 69 | 26 | **2.7× fewer** |

### English & code — not degraded

The efficiency gain on Indic scripts does **not** come at the cost of English or code:

- **English general:** 20 tokens (128K) vs 18 (old) — within 2 tokens, negligible.
- **English technical:** 19 vs 19 — exact parity.
- **Python:** 29 (128K) vs 33 (old) — Asterizer is **better**.
- **JavaScript:** 27 vs 27 — parity.

---

## Wider baseline comparison (8K smoke run, 55 probes)

The 8K build's fertility report benchmarks Asterizer against six industry tokenizers on `mean_bytes_per_token` over a 20 MB corpus sample (higher = better compression):

| Tokenizer | Mean bytes/token |
|---|---:|
| o200k_base | 4.73 |
| Gemma2 | 4.18 |
| Llama3 | 3.20 |
| Qwen2.5 | 3.15 |
| **Sarvam (Indic)** | 2.89 |
| **Asterizer** | 2.87 |
| cl100k_base | 2.69 |

Asterizer is competitive with the Indic-specialized **Sarvam** tokenizer — winning **34 of 55** head-to-head probes against it — while far exceeding `cl100k_base` on Indic text.

---

## Files

| Directory | Vocab | Contents |
|---|---|---|
| `128K/` | 128K | `tokenizer.json`, `tokenizer_config.json`, `special_tokens_map.json`, `fertility_report.json`, `manifest.json` |
| `64K/` | 64K | same bundle |
| `8K/` | 8K | same bundle |

### Tokenizer features

- **Class:** `PreTrainedTokenizerFast` (Hugging Face `tokenizers`)
- **Type:** BPE (byte-level), trained with `asterizer` v1.0 (seed 42)
- **Corpus:** 240 GB LUNA multilingual corpus (41.8M documents, 213.7 GB)
- **Special tokens:** `<|bos|>`, `<|endoftext|>`, `<|pad|>`, plus 29 chat/tool/FIM/multimodal tokens (`<|im_start|>`, `<|tool_call|>`, `<|fim_prefix|>`, `<|image|>`, etc.)
- **Chat template:** ChatML-style (`<|im_start|>` / `<|im_end|>`)
- **`model_max_length`:** 1,048,576

## Usage

```python
from transformers import AutoTokenizer

tok = AutoTokenizer.from_pretrained("ASTERIZER/LUNA-1B-Tokenizer", subfolder="128K")
ids = tok.encode("ಇದು ಕನ್ನಡ ಭಾಷೆಯ ಪರೀಕ್ಷಾ ವಾಕ್ಯ.")
print(tok.decode(ids))
```

## Related

- [LUNA-Tokenizer](https://huggingface.co/ASTERIZER/LUNA-Tokenizer) — original small tokenizer (100M/300M)
- [LUNA-1B Tokenizer Collection](https://huggingface.co/collections/ASTERIZER/luna-1b-tokenizer-6a9123d33ef778aa988fde53)
- [ASTERIZER](https://huggingface.co/ASTERIZER)

## License

`other` — the tokenizer artifacts derive from the LUNA multilingual corpus; verify upstream corpus terms before redistribution.