COILD-INDIC-MT — ICON 2026 Shared Task on Direct Indic-to-Indic Machine Translation

Motivation

Pivoting through English loses what makes these languages distinct.

Machine translation between Indian languages is, in practice, almost never between Indian languages. Where an Indic–Indic direction is supported at all, it is typically realised by pivoting through English or Hindi — lossy in exactly the places neighbouring Indian languages are rich: honorific agreement, classifier constructions, kinship terminology, and case and converb morphology. It also discards genuine areal proximity, and makes translation for a Bodo or Dogri speaker contingent on the state of English resources.

Why COIL-D

The COIL-D (Centre of Indian Language Data) project builds a unified repository of Indian language resources, sets benchmarking standards, and advances Machine Translation and NLP technologies for Human Language Technology applications. COILD-INDIC-MT is COIL-D's shared task focused exclusively on direct, non-pivoted translation between regional Indian languages.

A deliberately varied set

Rather than maximising language count, each pair isolates a different relationship: a cross-family pair convergent through contact (Assamese–Bodo); within-family pairs at differing degrees of relatedness (Bengali–Odia, Gujarati–Marathi, Kannada–Malayalam); a pair with an extremely low-resource source (Dogri–Punjabi); and a pair sharing a script family but diverging lexically (Urdu–Sindhi).

Shared task

12 directions, twelve languages

All twelve languages are listed in the Eighth Schedule of the Constitution of India. Participants may enter any subset of the six bidirectional language pairs.

#	Bidirectional pair	Family (src ↔ tgt)	Script (src ↔ tgt)
1	Assamese ↔ Bodo	Indo-Aryan ↔ Tibeto-Burman	Bengali–Assamese ↔ Devanagari
2	Kannada ↔ Malayalam	Dravidian ↔ Dravidian	Kannada ↔ Malayalam
3	Bengali ↔ Odia	Indo-Aryan ↔ Indo-Aryan	Bengali ↔ Odia
4	Gujarati ↔ Marathi	Indo-Aryan ↔ Indo-Aryan	Gujarati ↔ Devanagari
5	Dogri ↔ Punjabi	Indo-Aryan ↔ Indo-Aryan	Devanagari ↔ Gurmukhi
6	Urdu ↔ Sindhi	Indo-Aryan ↔ Indo-Aryan	Perso-Arabic ↔ Perso-Arabic

Objectives

Facilitate effective communication across India's diverse linguistic communities through direct machine translation, without relying on a pivot language.
Encourage multilingual and transfer learning across the community of Indian languages.
Provide a curated, high-quality multilingual parallel corpus to the community.
Help neighbouring communities communicate directly using machine translation.
Advance systems that preserve the linguistic and cultural diversity of close communities.
Establish a common evaluation platform for Indic-centric multilingual MT.

The corpus

~30,000 parallel sentences per direction

COILD-INDIC-MT provides a carefully curated, high-quality parallel corpus covering all six participating language pairs. The training and development sets are released in full, with the development set providing 500 source–target sentence pairs per translation direction for validation/tuning. The test set is released with source-side sentences only — its target references stay hidden for final evaluation.

Corpus splits — per language pair

Training set28,000

Development set Released
Source and target released — 500 sentence pairs per direction (e.g. 500 Assamese → Bodo + 500 Bodo → Assamese), for validation & tuning.1,000

Test set Hidden
Source given, target hidden — used for final evaluation.1,000

Total per pair~30,000

Across all six language pairs

Training set168,000

Development set Released
Source and target released — 500 sentence pairs in each of the 12 translation directions.6,000

Test set Hidden
Source given, target hidden — used for final evaluation.6,000

Total corpus~180,000

Access policy. The complete dataset (training, development, and test sets) is released only to registered participants, who must complete registration and agree to the dataset usage policy before obtaining access. Test set targets remain hidden — only source-side sentences are provided for that split, which is used for final evaluation. Development set targets are released in full.

Requesting the dataset. Once registered, the training, development, and test sets are distributed per language pair via a Hugging Face dataset request. Click a pair below to request access:

Use the same email address you registered with — only the team lead may submit the request, on behalf of the whole team.

Participation

Two submission tracks

Each participating team must submit at least one system to be eligible for submitting a system description paper. Teams may enter any subset of the six language pairs.

Track 1

Primary (Constrained)

Systems must be trained using only the datasets released by the organizers, and the model must be built from scratch.

Up to 3 runs per language pair, with one run designated as the primary submission.

Track 2

Open (Unconstrained)

Systems may utilize additional publicly available datasets or external resources and an open source model for finetuning.

Up to 3 runs per language pair, with one run designated as the primary submission.

Submitting your runs. Submit all runs, for either track, through the official submission form before the run-submission deadline listed in the Timeline.

Submission form ↗

All submissions must comply with the shared task guidelines and data usage policy specified by the organizers.

Timeline

Key dates

The gold marker tracks today's position on the line automatically, based on your device's date. All deadlines are 23:59 AoE (Anywhere on Earth) unless noted otherwise.

Updated 21 Oct 2026: the run submission deadline and all later dates have been extended. Previous dates are shown struck through.

22 Aug 2026