metadata

library_name: transformers
tags: []

Sarvam-2B

Sarvam-2B is a 2-billion parameter language model specifically optimized for Indian languages. It provides best in-class performance in 10 Indic languages (bn, gu, hi, kn, ml, mr, or, pa, ta, te) when compared with popular models like Gemma-2-2B and Llama-3.2-3B. It is also competitive against the much larger models like Llama-3.1-8B in these languages. More details can be found in our release blog.

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

Key Features

Optimized for 10 Indian Languages: Built from the ground up to support major Indian languages alongside English
Superior Token Efficiency: Achieves fertility rates of 1.4-2.1 across all supported languages, 2-4x more efficient than existing multilingual models
High-Quality Training Data: Trained on a curated corpus of ~4 trillion tokens with 2 trillion high-quality Indic tokens
Efficient Inference: 4-6x faster inference compared to larger models while matching or exceeding their performance on Indic language tasks

Model Architecture

Hidden size: 2048
Intermediate size: 11,008
Number of attention heads: 16
Number of hidden layers: 28
Number of key-value heads: 8
Maximum position embeddings: 8,192
Activation function: SwiGLU
Positional embeddings: Rotary (RoPE) with theta=10,000
Training: Grouped-query attention and bfloat16 mixed-precision

Performance

Academic Benchmarks (Zero-shot)

Flores English-to-Indic translation: 46.81 chrF++
CrossSum: 20.88 chrF++
XORQA: 26.47 F1
XQUAD: 41.58 F1

IndicGenBench (One-shot)

Flores English-to-Indic translation: 46.81% BLEU
CrossSum: 20.88% chrF++
XORQA: 26.47% F1
XQUAD: 41.58% F1

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

# Load model and tokenizer
model = AutoModelForCausalLM.from_pretrained("Sarvam/sarvam-2b")
tokenizer = AutoTokenizer.from_pretrained("Sarvam/sarvam-2b")

# Example usage
text = "कर्नाटक की राजधानी है:"
inputs = tokenizer(text, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=5)
result = tokenizer.decode(outputs[0])

Training Details

Training Infrastructure: Yotta's Shakti cluster
Hardware: 1,024 GPUs
Training Duration: 5 days
Framework: NVIDIA NeMo

License

Sarvam non-commercial license: See the LICENSE file

Acknowledgements

NVIDIA: for support with the NeMo codebase
Yotta: for sccess to the Shakti GPU cluster
AI4Bharat: for their academic partnership and expertise in Indian language technologies