LUNA-Tokenizer / README.md
REZIRETSA's picture
Apply CC BY-NC-ND 4.0 license
d35070a verified
|
Raw
History Blame Contribute Delete
1.87 kB
---
license: cc-by-nc-nd-4.0
language:
- en
pipeline_tag: token-classification
tags:
- gpt-neox
- tokenizer
- pythia
- luna
- text-generation
---
# LUNA Tokenizer
The **shared tokenizer** for the [LUNA model family](https://huggingface.co/ASTERIZER) — used by both [LUNA-100M](https://huggingface.co/ASTERIZER/LUNA-100M) and [LUNA-300M](https://huggingface.co/ASTERIZER/LUNA-300M).
It is the **EleutherAI / Pythia-160m (GPT-NeoX) tokenizer**, extracted into a standalone repository so both models — and any future LUNA variants — reference a single canonical source.
## Details
| Property | Value |
|---|---|
| **Tokenizer class** | `GPTNeoXTokenizer` |
| **Vocabulary size** | 50,304 (50,254 + EOS padding to multiple of 128) |
| **BOS token** | `<|endoftext|>` (id 0) |
| **EOS token** | `<|endoftext|>` (id 0) |
| **UNK token** | `<|endoftext|>` |
| **`add_prefix_space`** | false |
| **Upstream** | EleutherAI/Pythia-160m (gpt-neox-20b lineage) |
| **License** | Apache-2.0 (upstream EleutherAI tokenizer) |
## Files
| File | Description |
|---|---|
| `tokenizer.json` | Tokenizer vocabulary + merges (Hugging Face tokenizers format) |
| `tokenizer_config.json` | Tokenizer configuration (`GPTNeoXTokenizer`, special tokens) |
| `config.json` | Model-side config (`gpt_neox`, `vocab_size = 50304`) for reference |
## Usage
```python
from transformers import AutoTokenizer
tok = AutoTokenizer.from_pretrained("ASTERIZER/LUNA-Tokenizer")
ids = tok.encode("Hello, LUNA!")
print(tok.decode(ids))
```
## Related
- [LUNA-100M](https://huggingface.co/ASTERIZER/LUNA-100M)
- [LUNA-300M](https://huggingface.co/ASTERIZER/LUNA-300M)
- [LUNA-100M Program](https://huggingface.co/collections/ASTERIZER/luna-100m-program-6a9115383c18e52460bf67c9)
- [LUNA-300M Program](https://huggingface.co/collections/ASTERIZER/luna-300m-program-6a91195ab75671026f83ac94)