sarvam

The fully-trained version of this model is now available at https://huggingface.co/sarvamai/sarvam-1

Update (Aug 15, 2024): You can now get started with text completions and supervised finetuning using this notebook on Google colab!

This is an early checkpoint of sarvam-2b, a small, yet powerful language model pre-trained from scratch on 2 trillion tokens. It is trained to be good at 10 Indic languages + English. Officially, the Indic languages supported are: Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Oriya, Punjabi, Tamil, and Telugu.

The final checkpoint of sarvam-2b will be released soon, and it will be trained on a data mixture of 4 trillion tokens: containing equal parts English (2T) and Indic (2T) tokens.

The current checkpoint has not undergone any post-training. You can see the capabilities of the current checkpoint in this video.

The model was trained with NVIDIA NeMo™ Framework on the Yotta Shakti Cloud using HGX H100 systems.

Getting started:

from transformers import pipeline
pipe = pipeline(model='sarvamai/sarvam-2b-v0.5', device=0)
pipe('भारत के प्रथम प्रधानमंत्री', max_new_tokens=15, temperature=0.1, repetition_penalty=1.2)[0]['generated_text']
# 'भारत के प्रथम प्रधानमंत्री जवाहरलाल नेहरू थे।\n\n'

Tokenizer

sarvam-2b's tokenizer is built to be efficient for Indic languages and has an average fertility score of ~2 which is significantly lower than other models.

Here is a comparison of fertility scores between sarvam-2b and other popular models.

More technical details like evaluations and benchmarking will be posted soon.