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.
| Sarvam-2B | Llama-3.1 | Gemma-2 | GPT-4o | |
|---|---|---|---|---|
| ben_Beng | 2.07 | 8.02 | 3.72 | 2.34 |
| eng_Latn | 1.43 | 1.24 | 1.23 | 1.23 |
| guj_Gujr | 1.81 | 9.97 | 3.9 | 2.3 |
| hin_Deva | 1.4 | 2.67 | 1.96 | 1.65 |
| kan_Knda | 2.37 | 14.95 | 5.55 | 3.29 |
| mal_Mlym | 2.85 | 16.26 | 5.88 | 3.52 |
| mar_Deva | 1.77 | 3.99 | 3.2 | 2.56 |
| ory_Orya | 2.35 | 16.84 | 6.87 | 6.83 |
| pan_Guru | 1.68 | 8.19 | 3.37 | 2.72 |
| tam_Taml | 2.17 | 12.39 | 4.19 | 3.17 |
| tel_Telu | 2.14 | 13.3 | 4.57 | 3.06 |
| Average | 2.08 | 9.34 | 4.01 | 3.00 |
More technical details like evaluations and benchmarking will be posted soon.