---
language: ru
license: apache-2.0
tags:
  - tokenizer
  - unigram
  - russian
  - xnli
  - nlp-research
datasets:
  - facebook/xnli
---

# NIRVLab — Unigram Tokenizer for Russian XNLI

A **Unigram Language Model** tokenizer trained from scratch on the Russian (`ru`) subset
of the [facebook/xnli](https://huggingface.co/datasets/facebook/xnli) dataset.

## Training Details

| Parameter | Value |
|---|---|
| Algorithm | Unigram LM (SentencePiece-style) |
| Vocabulary size | 8,000 |
| Special tokens | `<s>, <pad>, </s>, <unk>, <mask>` |
| Corpus | `facebook/xnli` / `ru` — all splits |
| Corpus size | 800,404 sentences |
| Normalizer | Nmt + NFC Unicode |
| Pre-tokenizer | Metaspace (▁ prefix) |
| Shrinking factor | 0.75 |
| Max piece length | 16 |

## Evaluation Metrics

| Metric | Value |
|---|---|
| Tokens / char | `0.3315` |
| Fertility (tokens / word) | `2.2807` |
| Avg sequence length | `27.54` tokens |
| Vocabulary coverage | `1.0000` |

## Usage

```python
from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("NIRVLab/xnli-unigram-rusia")
tokens = tokenizer("Привет, мир!", return_tensors="pt")
```