Modelpublic

ManiKumarAdapala/Gemma3-En2Indic-NMT-270M

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

1likes29downloads

Model Card

Gemma-3-270M English → Indic Translator

A fine-tuned version of Gemma-3-270m-it for multilingual machine translation from English to 14 Indic languages.

This model is designed for lightweight, fast, and high-quality translation from English into major Indic languages while maintaining the conversational capabilities inherited from Gemma-3.

Model Details

Property	Value
Base Model	google/gemma-3-270m-it
Architecture	Gemma 3
Parameters	270 Million
Task	English → Indic Machine Translation
Framework	Hugging Face Transformers
Precision	bfloat16
Attention	Flash Attention 2
Generation	Beam Search
Beam Size	5
Sampling	Disabled (`do_sample=False`)
Padding Side	Left
Cache	Enabled (`use_cache=True`)

Supported Languages

Language	Language Code
Assamese	asm_Beng
Bengali	ben_Beng
Gujarati	guj_Gujr
Hindi	hin_Deva
Kannada	kan_Knda
Kashmiri	kas_Arab
Malayalam	mal_Mlym
Marathi	mar_Deva
Odia	ory_Orya
Punjabi	pan_Guru
Sanskrit	san_Deva
Tamil	tam_Taml
Telugu	tel_Telu
Urdu	urd_Arab

Input language is always English.

Dataset

The model was fine-tuned using the AI4Bharat BPCC (bpcc-seed-v2) multilingual parallel corpus.

The dataset contains parallel English–Indic sentence pairs covering multiple domains and language families.

Prompt Format

The model expects prompts in the following format:

text

Translate to {Target Language}:

{English Sentence}

Example:

text

Translate to Telugu:

Artificial Intelligence is changing healthcare.

Usage

python

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

MODEL_NAME = "ManiKumarAdapala/Gemma3-En2Indic-NMT-270M"

tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
tokenizer.padding_side = "left"

model = AutoModelForCausalLM.from_pretrained(
    MODEL_NAME,
    dtype=torch.bfloat16,
    attn_implementation="flash_attention_2",
    device_map={"": 0},
)

model.config.use_cache = True
model.eval()

EOT = tokenizer.convert_tokens_to_ids("<end_of_turn>")


@torch.inference_mode()
def translate(sentences, language):

    if isinstance(sentences, str):
        sentences = [sentences]

    prompts = [
        tokenizer.apply_chat_template(
            [
                {
                    "role": "user",
                    "content": f"Translate to {language}:\n\n{s}",
                }
            ],
            tokenize=False,
            add_generation_prompt=True,
        )
        for s in sentences
    ]

    inputs = tokenizer(
        prompts,
        return_tensors="pt",
        padding=True,
        add_special_tokens=False,
    ).to(model.device)

    outputs = model.generate(
        **inputs,
        max_new_tokens=256,
        do_sample=False,
        num_beams=5,
        eos_token_id=[
            tokenizer.eos_token_id,
            EOT,
        ],
        pad_token_id=tokenizer.pad_token_id,
    )

    generated = outputs[:, inputs["input_ids"].shape[1]:]

    return tokenizer.batch_decode(
        generated,
        skip_special_tokens=True,
    )


sentence = "Artificial Intelligence is transforming agriculture."

translation = translate(sentence, "Hindi")

print(translation[0])

Recommended Generation Settings

python

max_new_tokens = 256
num_beams = 5
do_sample = False
use_cache = True
padding_side = "left"
dtype = torch.bfloat16
attn_implementation = "flash_attention_2"

These settings are the same as those used in the provided inference notebook and are recommended for obtaining deterministic, high-quality translations.

Evaluation

The model was evaluated using sentence pairs from the BPCC dataset.

Metrics used:

—BLEU
—chrF2

Average benchmark results:

Language Code	Language	BLEU ↑	chrF2 ↑
asm_Beng	Assamese	44.1	48.2
ben_Beng	Bengali	55.3	54.3
guj_Gujr	Gujarati	54.1	53.5
hin_Deva	Hindi	61.9	59.5
kan_Knda	Kannada	41.0	46.6
kas_Arab	Kashmiri	8.4	31.6
mal_Mlym	Malayalam	45.3	46.7
mar_Deva	Marathi	50.7	50.6
ory_Orya	Odia	40.6	44.3
pan_Guru	Punjabi	48.7	49.2
san_Deva	Sanskrit	38.4	41.2
tam_Taml	Tamil	48.9	49.2
tel_Telu	Telugu	49.8	49.7
urd_Arab	Urdu	33.7	55.2
Average	14 Languages	44.4	48.6

Known Limitations

Like most compact multilingual translation models, this model has a few limitations.

—Numerical values may occasionally change during translation.
—Rarely, Latin characters may appear within Indic script outputs.
—Translation quality varies across languages, with lower-resource languages generally being more challenging.
—Not intended for legal, medical, or other safety-critical translation tasks without human verification.

Citation

bibtex

@misc{Gemma3-En2Indic-NMT-270M,
  title = {Gemma3-En2Indic-NMT-270M: English to Indic Neural Machine Translation},
  author = {Adapala, Mani Kumar},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/ManiKumarAdapala/Gemma3-En2Indic-NMT-270M}
}

Acknowledgements

This work builds upon:

—Google for the Gemma-3 model.
—AI4Bharat for the BPCC multilingual parallel corpus.
—Hugging Face for the Transformers ecosystem.
—The open-source community for tools and libraries that enabled this work.