Machine translation (MT) for Indian languages remains constrained by the limited availability of high-quality, Indic-centric parallel corpora and evaluation benchmarks. Existing multilingual resources are largely constructed from English-pivot content and often fail to capture the linguistic diversity, cultural complexity, and domain-specific characteristics of Indian languages. We present COILD, an Indic-centric parallel corpus comprising over 1.16 million human-translated and human-verified sentence pairs, covering 20 Indian language pairs across the Indo-Aryan, Dravidian, Tibeto-Burman, and Austro-Asiatic language families. The corpus is built entirely from original Indian language sources collected from licensed repositories spanning eight domains with direct real-world applicability. Furthermore, we introduce a domain-centric benchmark comprising 2,000 expert-verified sentences to enable consistent multilingual and cross-lingual evaluation across Indian language pairs. To validate the effectiveness of COILD, we fine-tune two representative multilingual neural machine translation models, IndicTrans2-Distilled and NLLB-200. Experimental results demonstrate consistent improvements across language pairs, domains, automatic evaluation metrics, and human evaluation, highlighting the effectiveness of high-quality Indic-centric supervision. COILD provides a valuable training and evaluation resource for advancing multilingual machine translation and future multilingual language models for Indian languages.
Figures & tables
Figure 1: Overview of the COILD framework and evaluation pipeline. The workflow consists of three stages: (1) dataset construction , using a dual-centric and bidirectional strategy with Hindi as the Indo-Aryan pivot and Tamil as the Dravidian pivot, resulting in 20 language pairs and 40 translation directions; (2) model fine-tuning , where IndicTrans2 and NLLB-200 are fine-tuned on the COILD parallel corpus; and (3) evaluation , comprising automatic evaluation using five MT metrics, domain-wise evaluation across eight application domains, and human evaluation based on translation adequacy, fluency, and overall quality using a 5-point Likert scale.
Source
Target
Sentence Pairs
Token Count
Vocabulary
Source
Target
Source
Target
Hindi
Assamese
94,872
19,86,178
15,90,680
76,956
1,18,170
Hindi
Bengali
73,066
14,89,733
11,76,804
58,937
79,616
Hindi
Bodo
47,599
9,88,036
7,62,802
47,698
74,469
Hindi
Dogri
58,896
11,93,781
12,09,040
53,385
56,814
Hindi
Gujarati
96,612
20,22,924
16,84,920
78,395
1,15,035
Table 1: Statistics of the COILD corpus. The corpus contains Hindi-centric (Hindi → 17 target languages) and Tamil-centric (Tamil → 3 target languages) parallel corpora. We report the number of aligned sentence pairs, total token counts, and vocabulary sizes for the source and target sides. The COILD corpus contains approximately 1.16 million human-translated sentence pairs, comprising 44.79 million tokens across the source and target sides.
Figure 2: Domain-wise distribution of the COILD benchmark, comprising 2,000 Hindi source sentences across eight domains. The Education domain contains 600 sentences, while each of the remaining seven domains contains 200 sentences.
Hyperparameter
IndicTrans2
NLLB-200
Epochs
5
5
Learning Rate
2×10−5
2×10−5
LR Scheduler
Cosine
Linear
Warm-up
1%
3%
Optimizer
AdamW (8-bit)
AdamW (8-bit)
Effective Batch Size
64
64
Table 2: Hyperparameter configuration used for fine-tuning IndicTrans2 and NLLB
Figure 3: Overall translation performance after fine-tuning on COILD for IndicTrans2-Distilled (top) and NLLB-200 (bottom). Heatmaps show the improvement over the corresponding pretrained baseline ( Δ = Fine-tuned − Baseline) across five evaluation metrics: BLEU, chrF++, TER, BERTScore, and COMET. Columns represent translation directions, rows denote evaluation metrics, and larger positive values indicate better performance achieved through fine-tuning on the COILD corpus.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
IndicTrans2
NLLB
Direction
Adequacy
Fluency
Adequacy
Fluency
Baseline
Finetune
Δ
Baseline
Finetune
Δ
Baseline
Finetune
Δ
Baseline
Finetune
Δ
Assamese → Hindi
2.16
2.49
0.33
3.01
3.35
0.34
2.52
2.71
0.19
3.42
3.54
0.12
Hindi → Assamese
2.19
2.12
-0.07
3.03
2.82
-0.21
2.02
2.32
0.30
2.66
2.80
0.14
Bengali → Hindi
3.86
4.22
0.36
4.15
4.44
0.29
3.43
3.89
0.47
3.52
4.02
0.49
Hindi → Bengali
3.71
4.09
0.38
3.99
4.27
0.28
3.50
3.85
0.35
3.56
4.02
0.46
Appendix
Table 3: Human evaluation results comparing the baseline and fine-tuned versions of IndicTrans2 and NLLB. Adequacy and Fluency were rated by human annotators on a 5-point Likert scale (1 = Worst, 2 = Poor, 3 = Average, 4 = Good, 5 = Excellent). Positive Δ values indicate improvements after fine-tuning and are colour-coded to highlight the magnitude of performance changes after fine-tuning. Green shades indicate improvements, red shades indicate degradations. (Note: Odia Language human evaluation is underway.)
Figure 4: Per-domain change in BLEU after finetuning on COILD, for IndicTrans2 (top) and NLLB-200 (bottom). Columns are translation directions, rows are benchmark domains; cell values are Δ BLEU (finetune − baseline).
Figure 5: Plot shows the overall baseline and finetune performance on the COILD Benchmark dataset for IndicTrans2.
Figure 6: Plot shows the overall baseline and finetune performance on the COILD Benchmark dataset for NLLB.
We study English-to-Prakrit machine translation in a low-resource setting where the target language is unsupported by IndicTrans2. We adapt the multilingual model by mapping Prakrit to the Hindi language tag (hin_Deva) without modifying the tokenizer, vocabulary, or architecture. Using a 1,474-pair Maharashtri Prakrit parallel corpus and evaluation on a 20-sample Ardhamagadhi test set, we report corpus BLEU improvements over an untuned baseline. The results indicate that script-compatible language routing can enable feasible transfer to unsupported classical languages, while highlighting limitations due to data scarcity and dialect mismatch. Our code and trained models are released to the public for further exploration https://github.com/D3v1s0m/indictrans2-prakrit-mt.
Om Choksi, Smit Kareliya, Shrikant Malviya +1
Sardar Vallabhbhai National Institute of Technology, Surat
Large Language Models (LLMs) have transformed conversational AI, yet high-quality multilingual code-mixed dialogue resources remain scarce, particularly for Indic languages where speakers naturally alternate between English and their native language in both native-script and Romanized forms. We present IndicTalk, one of the largest multilingual Indic code-mixed conversational corpora, comprising over 13,28,604 event-grounded multi-turn conversations across 18 language varieties covering 9 Indic languages. The corpus is generated through a fully automated pipeline that combines real-world news grounding, persona-conditioned dialogue generation using multilingual LLMs, and automatic quality validation. Extensive linguistic, automatic, and human evaluations demonstrate that IndicTalk produces fluent, coherent, and naturally code-mixed conversations across both script variants. We will release IndicTalk to support the development and evaluation of multilingual conversational AI for underrepresented Indic languages. The dataset is available at: https://huggingface.co/datasets/LingoIITGN/IndicTalk .
Open web-scale pre-training corpora remain concentrated in English, limiting multilingual LLM development. We introduce MultiSynt/MT, an open synthetic parallel corpus with approximately 4.8 trillion target-language tokens across 36 languages, produced by translating 100 billion high-quality Nemotron-CC tokens with Tower+ and OPUS-MT/HPLT-MT systems. For many medium- and lower-resource European languages, this is the largest openly available pre-training resource. Across five high- and medium-resource languages, reference LLMs trained on MultiSynt/MT reach the final score of HPLT 2.0, a native-data baseline, using roughly 72% fewer pre-training tokens, and outperform it by approximately 15% relative at a matched 100B-token training budget. Our analyses also identify evaluation blind spots: standard multiple-choice benchmarks miss translation-quality differences that a fluency-sensitive LLM-as-judge protocol recovers on the trained LLMs without detecting a deficit relative to its native-data baseline, while Norwegian idiomatic and culturally grounded tasks remain better served by native data. We release the corpus, including row-aligned translations from multiple systems, to support controlled research on multilingual pre-training data and evaluation.
Maximilian Idahl, Jörg Tiedemann, Sampo Pyysalo +19
ellamind · Leibniz University Hannover · University of Helsinki +14