IndicTrans2

What it does

Datasets

The project releases the BPCC parallel corpus, BPCC back-translation data, and IN22 evaluation benchmarks.checked 4 Oct 2026

Deployment

The README documents local command-line and Python inference workflows; it does not describe a hosted API or commercial integrations.checked 4 Oct 2026

Evaluation

The project reports chrF++, BLEU, and COMET metrics and includes evaluation scripts and benchmarks.checked 4 Oct 2026

Inference

Users can run batch or paragraph translation through Python interfaces using Fairseq or CTranslate2 models.checked 4 Oct 2026

Installation requirements

The README recommends Python 3.7 or later and notes that some data preparation scripts require SentencePiece and GNU parallel.checked 4 Oct 2026

Intended users

The repository describes its open data, models, and scripts as supporting use and adoption within the research community.checked 4 Oct 2026

Language coverage

The models support English and Indic languages, including multiple scripts for Kashmiri, Manipuri, and Sindhi.checked 4 Oct 2026

License

The repository code is under the MIT License, while released datasets have artifact-specific licenses including CC0 and CC-BY-4.0.checked 4 Oct 2026

Long context

RoPE-based model variants are described as supporting sequence lengths up to 2048 tokens.checked 4 Oct 2026

Model variants

The repository provides base and distilled model checkpoints, including Fairseq and Hugging Face artifacts.checked 4 Oct 2026

Purpose

IndicTrans2 is an open-source multilingual neural machine translation model for the 22 scheduled Indic languages.checked 4 Oct 2026

Security and compliance

The opened project pages do not state security certifications or compliance assurances.checked 4 Oct 2026

Training

The repository includes scripts and instructions to train or fine-tune multilingual translation models.checked 4 Oct 2026

Translation directions

Released model families cover English-to-Indic, Indic-to-English, and Indic-to-Indic translation.checked 4 Oct 2026