Update README.md

README.md

@@ -1213,4 +1213,85 @@ configs:
   data_files:
   - split: train
     path: vi/train-*
+language:
+- af
+- ar
+- bg
+- ca
+- cs
+- da
+- de
+- el
+- en
+- es
+- et
+- eu
+- fa
+- fi
+- fr
+- ga
+- he
+- hi
+- hr
+- hu
+- hy
+- id
+- it
+- ja
+- ko
+- lt
+- lv
+- mr
+- nl
+- no
+- pl
+- pt
+- ro
+- ru
+- sa
+- sk
+- sl
+- sr
+- sv
+- ta
+- te
+- tr
+- uk
+- ur
+- vi
 ---
+
+# Multilingual Tokenizer Benchmark 
+
+This dataset includes pre-processed wikipedia data for tokenizer evaluation in 45 languages. 
+
+## Usage
+The dataset allows us to easily calculate tokenizer fertility and the proportion of continued words on any of the supported languages. In the example below we take the Mistral tokenizer and evaluate its performance on Slovak. 
+
+```python 
+from transformers import AutoTokenizer
+from datasets import load_dataset
+import numpy as np
+
+def calculate_metrics(tokens):
+    tmp = np.array([len(y) for y in tokens])
+    return {'fertility': np.mean(tmp), 'cont_prop': np.count_nonzero(tmp > 1) / tmp.shape[0]}
+
+tokenizer_name = 'mistralai/Mistral-7B-v0.1'
+language = 'sk' #Slovak
+tokenizer = AutoTokenizer.from_pretrained(tokenizer_name)
+ds = load_dataset('occiglot/tokenizer-wiki-bench', name=language, split='clean')
+
+remove_columns = list(set(ds.column_names) - set(["text"]))
+ds = ds.map(lambda x: {'tokens': tokenizer(x['split_text'], add_special_tokens=False)['input_ids']} ,num_proc=256, remove_columns=remove_columns, batched=False)
+remove_columns = None#list(set(ds.column_names))
+ds = ds.map(lambda x: calculate_metrics(x['tokens']), num_proc=256, remove_columns=remove_columns,  batched=False)
+df = ds.to_pandas()
+
+print('Fertility: ', df.fertility.mean())
+print('Prop. continued words:', df.cont_prop.mean())
+```
+
+## Dataset Creation
+
+We loosely follow the approach of [Rust _et al.](https://arxiv.org/abs/2012.15613) using the fast [UDPipe](https://ufal.mff.cuni.cz/udpipe) to pre-split documents into words and subsequently run the tokenizer over isolated words. For all languages we use the respective November 2023 snapshot from [Wikipedia](wikimedia/wikipedia). Since Wikipedia, by nature, contains significantly more numbers and dates than other text and most tokenizers split those into single digits, we filtered all lone-standing numbers from the documents. Additionally, we removed any documents that still contained non-parsed HTML code (less than 1%).