A fine-tuned translation model supporting 55 Indian languages including 33 low-resource languages. Translates bidirectionally between English and any supported Indic language.

Key Features

55 Indian languages — covers both major and low-resource regional languages
Bidirectional — English to Indic and Indic to English
1.3B parameter encoder-decoder model, ready to use out of the box
Fast inference — supports CTranslate2 int8 quantization for production deployments
Extended vocabulary — 256,243 tokens with 33 newly added language codes

Supported Languages

Language	Code	Script	Language	Code	Script
Ahirani	ahr_Deva	Devanagari	Kumaoni	kfy_Deva	Devanagari
Assamese	asm_Beng	Bengali	Kurukh	kru_Deva	Devanagari
Awadhi	awa_Deva	Devanagari	Kutchi	kfr_Deva	Devanagari
Bagheli	bfy_Deva	Devanagari	Magahi	mag_Deva	Devanagari
Bagri	bgq_Deva	Devanagari	Maithili	mai_Deva	Devanagari
Banjari	brj_Deva	Devanagari	Malayalam	mal_Mlym	Malayalam
Bengali	ben_Beng	Bengali	Manipuri	mni_Beng	Bengali
Bhili	bhb_Deva	Devanagari	Manipuri	mni_Mtei	Meitei Mayek
Bhojpuri	bho_Deva	Devanagari	Marathi	mar_Deva	Devanagari
Bodo	brx_Deva	Devanagari	Marwari	mwr_Deva	Devanagari
Braj Bhasha	bra_Deva	Devanagari	Mewari	mtr_Deva	Devanagari
Bundeli	bns_Deva	Devanagari	Mizo	lus_Latn	Latin
Chhattisgarhi	hne_Deva	Devanagari	Nepali	npi_Deva	Devanagari
Dakhini	dcc_Deva	Devanagari	Odia	ory_Orya	Odia
Dogri	doi_Deva	Devanagari	Pahari	phr_Deva	Devanagari
Garhwali	gbm_Deva	Devanagari	Punjabi	pan_Guru	Gurmukhi
Garo	grt_Latn	Latin	Rajasthani	raj_Deva	Devanagari
Gondi	gon_Deva	Devanagari	Sambalpuri	spv_Orya	Odia
Gujarati	guj_Gujr	Gujarati	Sanskrit	san_Deva	Devanagari
Haryanvi	bgc_Deva	Devanagari	Santali	sat_Olck	Ol Chiki
Hindi	hin_Deva	Devanagari	Sindhi	snd_Arab	Arabic
Ho	hoc_Deva	Devanagari	Sora	srb_Latn	Latin
Kangri	xnr_Deva	Devanagari	Tamil	tam_Taml	Tamil
Kannada	kan_Knda	Kannada	Telugu	tel_Telu	Telugu
Kashmiri	kas_Arab	Arabic	Tulu	tcy_Knda	Kannada
Kashmiri	kas_Deva	Devanagari	Urdu	urd_Arab	Arabic
Khasi	kha_Latn	Latin	Wagdi	wbr_Deva	Devanagari
Khortha	kho_Deva	Devanagari
Kodava	kfa_Knda	Kannada
Konkani	kok_Deva	Devanagari

How to Use

Quick Start (Transformers)

python

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

# Load model
model = AutoModelForSeq2SeqLM.from_pretrained("Anonym-050326/nirukti-translate-1.3b")
tokenizer = AutoTokenizer.from_pretrained("Anonym-050326/nirukti-translate-1.3b")

# Translate English to Hindi
tokenizer.src_lang = "eng_Latn"
inputs = tokenizer("Hello, how are you?", return_tensors="pt")
translated = model.generate(
    **inputs,
    forced_bos_token_id=tokenizer.convert_tokens_to_ids("hin_Deva"),
    max_new_tokens=128,
)
print(tokenizer.decode(translated[0], skip_special_tokens=True))

CTranslate2 (Fastest — Recommended for Production)

python

import ctranslate2
from transformers import AutoTokenizer

translator = ctranslate2.Translator("ct2-int8", device="cuda", compute_type="int8_float16")
tokenizer = AutoTokenizer.from_pretrained("Anonym-050326/nirukti-translate-1.3b")
tokenizer.src_lang = "eng_Latn"

text = "Hello, how are you?"
encoded = tokenizer(text, return_tensors=None, max_length=256, truncation=True)
tokens = tokenizer.convert_ids_to_tokens(encoded["input_ids"])

result = translator.translate_batch([tokens], target_prefix=[["hin_Deva"]], beam_size=5)
output_tokens = result[0].hypotheses[0][1:]  # skip language token
output_ids = tokenizer.convert_tokens_to_ids(output_tokens)
print(tokenizer.decode(output_ids, skip_special_tokens=True))

Translating Between Any Language Pair

Set tokenizer.src_lang to the source language code and use the target code as forced_bos_token_id (Transformers) or target_prefix (CTranslate2).

python

# Tamil to English
tokenizer.src_lang = "tam_Taml"
# forced_bos_token_id=tokenizer.convert_tokens_to_ids("eng_Latn")

# English to Marwari
tokenizer.src_lang = "eng_Latn"
# forced_bos_token_id=tokenizer.convert_tokens_to_ids("mwr_Deva")

Testing Scripts

This repository includes evaluation and testing scripts in the testing_scripts/ directory:

Script	Description
`testing_scripts/run_evaluation.py`	Multi-GPU evaluation runner (BLEU + chrF++)
`testing_scripts/download_datasets.py`	Download standard Indic translation benchmarks
`testing_scripts/excel_to_csv.py`	Convert Excel test datasets to evaluation format

Running Evaluation

bash

# 1. Convert to CTranslate2 int8 for fast inference
ct2-opennmt-py-converter --model_path Anonym-050326/nirukti-translate-1.3b --output_dir ct2-int8 --quantization int8

# 2. Download benchmark datasets
python testing_scripts/download_datasets.py --output-dir eval_data

# 3. Run multi-GPU evaluation
python testing_scripts/run_evaluation.py \
    --model ct2-int8 \
    --tokenizer Anonym-050326/nirukti-translate-1.3b \
    --manifest eval_data/manifest.json \
    --output-dir results \
    --num-gpus 4 --batch-size 64 --beam-size 5

Training Details

Architecture

Base: 1.3B parameter encoder-decoder translation model
Fine-tuning: LoRA (rank 32, alpha 64) on attention + FFN layers, merged into base weights
Trainable parameters: ~0.75% of total during training
2-stage training: embedding divergence followed by LoRA fine-tuning

Hyperparameters

Parameter	Value
Learning rate	3e-5 (cosine schedule)
Warmup steps	1,000
Effective batch size	256
Epochs	5
Max sequence length	128 tokens
Precision	BF16
Label smoothing	0.1
Language balancing	Temperature sampling (T=5.0)

Training Data

Trained on a curated combination of parallel corpora covering all 55 languages, with hash-based deduplication and quality filtering. Low-resource language pairs are upsampled using temperature-based balancing to ensure adequate representation.

Compute

Training: 2x NVIDIA A100-SXM4-80GB, PyTorch DDP
Inference: CTranslate2 int8 quantization

Limitations

Low-resource languages (33 newly added) have less training data and lower translation quality than high-resource languages
Performance varies by script — languages with multiple script variants may show different quality levels
Training data is primarily from news, government, and religious domains; informal or conversational text may translate differently
Very short sentences (1-5 words) are harder for the model due to limited context