Tokenizer Card for Ansh-128k!

The tokenizer model Ansh-128k - is trained on a dataset of 22 Official Indic languages and English. We propose the name Ansh as this tokenizer is designed to meticulously identify every essential token (Ansh in Sanskrit) of our diverse Indic languages. This model is the advanced version of the Ansh-160k which was trained on 18 Indic languages and English.

Model Description

India is a vast country that has a multi-lingual culture that covers 22 Official languages and more than 1700 languages and dialects. It has been observed that various languages share words among themselves, sometimes even across language families. To capitalize on this observation, we trained our tokenization model with a vocabulary size of 128,000 (128k) using the dataset of Wikipedia articles and Sangraha dataset in 22 Indic languages and English by applying the Byte-Pair Encoding (BPE) algorithm. When compared among all the popular open-source tokenizers trained on multilingual Indic languages on fertility scores, our model outperformed them in 9 Indic languages.

• Developed by: Lingo Research Group at IIT Gandhinagar
• Language(s) (NLP): Multilingual (22 Indic Languages and English)
• License: Apache 2.0

How to Get Started with the Model 👨🏻‍💻

Use the code below to get started with the model.

from transformers import AutoTokenizer
try:
    tokenizer = tokenizer = AutoTokenizer.from_pretrained("LingoIITGN/Ansh-128k"))
    print("Tokenizer loaded successfully!")
except Exception as e:
    print(f"Error loading tokenizer: {e}")
    print("Please ensure you have the correct model name and are connected to the internet.")
    exit()
 
input_text = "Hello, world! This is an example of how to use the tokenizer."
#input_text = 'मुझे यह presentation कल morning तक submit करना है। '
#input_text = 'What is capital city of India?'

encoded_input = tokenizer.encode(example_text)
print("\nOriginal Text:", example_text)
print("Encoded (Token IDs):", encoded_input)

decoded_output = tokenizer.decode(encoded_input)
print("Decoded Text:", decoded_output)

Evaluation

[More Information Needed]

Results 🏆

Comparison of Fertility Scores among popular open-source tokenizers trained on multilingual Indic languages and Ansh-128k tokenizers across the 22 Indic languages and English.

Tokenizers Results

Language	Ansh-128k	MuRIL	IndicBERTv2	Ansh-160k	Llama-3.1	NLLB	XLMRoBERTa	Gemma	Sarvam-1
Tamil	1.915	1.844	1.790	1.899	11.941	2.742	2.486	2.524	2.590
Kannada	1.909	1.953	1.815	1.862	14.239	2.846	2.507	3.349	2.654
Malayalam	2.210	2.337	2.177	2.236	16.064	3.406	2.968	3.612	3.363
Maithili	1.474	1.832	1.695	1.561	3.246	1.955	2.133	2.152	2.503
Konkani	1.941	2.491	2.221	2.072	4.037	2.617	2.581	2.727	2.992
Telugu	1.940	2.069	1.873	2.010	13.240	2.859	2.552	3.143	2.693
Odia	1.546	1.714	1.539	1.587	15.535	2.149	2.196	4.523	2.494
Bengali	1.542	1.442	1.461	1.509	8.200	2.205	2.140	1.767	2.045
Nepali	1.376	1.413	1.411	1.428	3.611	1.898	1.643	2.027	2.358
Punjabi	1.415	1.420	1.341	1.434	7.855	1.843	1.798	2.789	1.726
Urdu	1.285	1.314	1.393	1.270	3.003	1.589	1.430	1.687	8.417
Hindi	1.245	1.276	1.272	1.246	2.757	1.546	1.525	1.442	1.480
Gujarati	1.537	1.587	1.459	1.495	9.651	2.145	2.062	2.358	2.093
Kashmiri	1.540	2.131	2.646	1.619	4.026	2.849	2.985	3.053	9.248
Marathi	1.585	1.579	1.521	1.573	4.010	2.207	2.011	2.012	1.979
Sindhi	1.300	1.354	1.630	1.333	2.938	1.621	1.532	2.101	8.165
Assamese	1.662	1.770	1.686	1.724	8.051	2.191	2.875	2.728	4.334
Sanskrit	2.444	2.855	2.732	2.470	5.034	3.453	3.344	3.562	3.949
Bodo	1.486	2.761	1.886	2.499	3.855	3.008	3.068	3.057	3.136
Santhali	1.333	1.144	1.966	4.538	13.456	2.994	2.095	5.634	14.402
Dogri	1.539	1.512	1.457	1.525	2.810	1.721	1.717	1.658	1.789
Manipuri	4.416	1.436	2.497	4.407	13.184	2.237	2.326	9.272	13.496
English	1.545	1.368	1.373	1.449	1.384	1.480	1.470	1.415	1.743
Overall	1.641	1.899	1.893	2.348	6.024	2.498	2.439	3.123	5.963

Model Card Contact ✉️

Lingo Research Group at IIT Gandhinagar, India
Mail at: lingo@iitgn.ac.in