IndicTrans2: Towards High-Quality and Accessible Machine Translation Models for all 22 Scheduled Indian Languages

Training datasets

• ~230M bitext pairs, including ~126M newly added, make BPCC the largest publicly available Indic parallel corpus; its 644K human-translated En-X seed spans all 22 languages.The seed includes seven languages for which no manually created parallel data existed before this work.Source: §1 · p. 2
• BPCC supplements existing filtered bitext with newly mined pairs from monolingual and document-aligned sources, plus ~800 million back-translated sentences.New sources include archive.org, IndicCorp v2, NPTEL, UGCResources, and Prabhupada Vani; existing Samanantar and NLLB pairs were filtered using LaBSE cosine-similarity thresholds.Source: §1 · pp. 2–3

Benchmarks

• IN22 is the first n-way parallel benchmark for all 22 languages, with source-original English and India-specific IN22-Gen and everyday-conversation IN22-Conv subsets.IN22-Gen covers 13 domains, while IN22-Conv covers 16 conversation domains.Source: §1 · p. 3

Open Access

• The authors release IndicTrans2 derivatives, training data, source code, and tools openly; IndicTrans2-Dist is about one-fifth the size of IndicTrans2 with comparable performance.IndicTrans2-Dist is approximately 211M in size.Source: §1 · p. 3

Models

• IndicTrans2 is the first model supporting all 22 scheduled languages, performs at par with NLLB and Google and Microsoft models, and its M2M variant covers 462 directions.The paper reports improved English-to-Indic chrF++ translation quality across languages with IndicTrans2.Source: §1 · p. 3, Figure 1 · p. 4

Languages of India

• The 22 scheduled languages span four major language families and up to 12 scripts, while many still lack online presence and high-quality NLP technologies.The passage also notes long-term linguistic similarities arising from language relatedness and contact.Source: §2 · pp. 4–5

Datasets

• Indic parallel corpora grew from small annotated collections to mined datasets, but existing efforts still omitted some scheduled languages and prompted seed-data creation for low-resource languages.ILCI created 50K sentences per language for 12 languages; Samanantar contained 49.7 million pairs across 11 Indic languages (Ramesh et al., 2022), while NLLB had 448 million English-centric pairs covering 19 Indic languages.Source: §2 · pp. 5–6

Benchmarks and Shared Tasks

• Indic MT evaluation expanded through shared tasks and multilingual benchmarks, whose language coverage remained uneven across resources.WMT began with Hindi-English, and WAT expanded Indic coverage over successive years (Nakazawa et al., 2018; 2020; 2021a; 2022). FLORES 101 included 14 Indic languages (Goyal et al., 2022), NLLB 19, and NTREX 16.Source: §2 · pp. 6–7

Neural MT models

• Indic NMT progressed from recurrent to Transformer models, while attention, subword modeling, and multilingual transfer supported fluent translation and low-resource language performance.Larger corpora and models improved translation quality (Ramesh et al., 2022), and recent work explored extremely low-resource languages (Bapna et al., 2022).Source: §2 · p. 7

3.2 Building the IN22 Test set

• IN22 addresses limited existing coverage by spanning all 22 Indic languages and emphasizing diverse domains, realistic sentence lengths, and India-relevant topics.The stated gaps include sparse representation of Indian names, locales, and culture-specific terms in general-purpose benchmarks.Source: §3.2 · p. 8

3.2.1 Corpus Description

• IN22 combines 512-sentence Wikipedia and 512-sentence web-source subsets with a 1,503-sentence conversation subset for sentence-, turn-, and document-level evaluation.Wikipedia and web sources cover 13 domains, while conversations cover 16; the web subset draws on non-crawlable sources and focuses more on India-centric content.Source: §3.2.1 · pp. 8–9
• IN22 contains 2,527 original English sentences translated into 22 Indic languages and supports evaluation in 500+ directions, including 462 inter-Indic directions.It includes source domains and unedited context sentences as metadata for fine-grained analysis.Source: §3.2.1 · p. 9

3.2.2 Source Selection

• IN22 source sentences come from Wikipedia categories, Indian government websites, digital libraries, and conversations created from predefined prompts and scenarios.Wikipedia sentences under six or over 80 words and those with four-gram overlap with FLORES-200 were filtered; context sentences were also assembled.Source: §3.2.2 · p. 9

3.2.3 Quality Control Procedure.

• For IN22 quality control, Shoonya disables MT outputs and translations are accepted when MT-system BLEU-score differences fall within δ = 10.When overlap is high, a new annotator is assigned and quality control is repeated.Source: §3.2.3 · p. 10

3.3 Building the BPCC Training Set

• BPCC-H was created as a human-translated seed corpus because mining alone is less suited to very-low-resource languages, and seed data can support MT and other NLP tools.The motivation draws on prior seed-corpus efforts including ILCI, ALT, and NLLB-Seed.Source: §3.3 · pp. 10–11

3.3.1 Corpus Description

• BPCC-H-Wiki contains 644.3K manually translated sentence pairs, including more than 10K pairs each for Bodo, Dogri, Santali, and Goan Konkani.BPCC-H-Daily adds 139.7K bitext pairs across 21 Indic languages except Sindhi; the released seed corpora are not yet n-way parallel because language teams translated different English batches.Source: §3.3 · p. 10, §3.3.1 · p. 11

3.3.2 Translation Details

• BPCC-H translates English source sentences into Indic languages, using Wikipedia selection for domain coverage and three-sentence blocks to provide translation context.Translators could post-edit existing MT outputs where feasible, and the English-original design entails translationese on the Indic side.Source: §3.3.2 · p. 11

4 Mining Training Data at Scale

• Mining added ~126M sentence pairs, bringing deduplicated corpora to ~230.5M (~5× Ramesh et al. (2022)); monolingual and comparable-corpus mining contributed 121M and 4.35M pairs, respectively.Comparable-corpus mining produced a diverse set spanning topics including Religion, Education, and Legal content.Source: §4 · p. 12

4.1 Mining from Monolingual Corpora

• Monolingual mining yielded 121M bitext pairs using a larger collection of Indic and English sentences than Samanantar.The pipeline used LaBSE, whose cosine similarities correlated more strongly with human STS ratings than LASER3 on a macro scale, falling back to LASER3 for unsupported languages. It curated and deduplicated sentence data, then searched English sentence embeddings for matches above a 0.80 cosine threshold; low-resource languages instead relied on human-translated data.Source: §4 · p. 12, §4.1 · pp. 13–15

4.2 Mining from Comparable Corpora

• Comparable-corpus mining extracted 4.5M sentence pairs across 17 Indic languages from high-quality sources spanning domains such as Education, Legal, and Religion.It aligned documents, used margin-based mining in both directions and retained intersecting matches, with a 1.06 margin threshold, 4 nearest neighbors, and a 0.80 cosine threshold for high-resource languages.Source: §4.2 · pp. 15–16

4.3 Filtering Existing Mined Parallel Corpora

• Quality filtering reduced Allen-NLLB from 448.1M to 104.2M sentence pairs, dropping close to 76% of the data.The authors also filtered Samanantar, reducing it from 49.7M to 19.4M pairs; human evaluation confirmed that most Santali-English data in Allen-NLLB was noisy.Source: §4 · p. 12, §4.3 · p. 17

5.1 Training Data

• 230.5 million sentence pairs, including 2.2 million professionally annotated gold pairs, form the training data drawn from mined, human-annotated, filtered, and additional seed corpora.Additional seed data from NLLB-seed, ILCI, and MASSIVE totals approximately 1.47 million sentence pairs; ILCI contributes around 1.34 million across 16 languages, and manually translated MASSIVE Indic data contributes 139,000 across seven languages.Source: §5.1 · p. 17

5.2 Preprocessing

• Preprocessing removes benchmark-overlapping bitext, maps Indic numerals to English numerals, tags matching special-case spans, and converts many Indic languages to Devanagari where feasible.Deduplication leaves approximately 230.5M bitext pairs; special tags mark emails, URLs, dates, numbers, and other spans appearing in both source and target, while script unification yields five scripts overall.Source: §5.2 · p. 18

5.3 Tokenization

• Separate SentencePiece BPE tokenizers use 32K English and 128K Indic vocabularies, with low-resource languages upsampled and high-resource languages capped at 3M sentences each.The tokenizers are trained on sampled monolingual IndicCorp v2 and NLLB data and include <dnt> and </dnt> tags; source and target language tags specify translation direction.Source: §5.3 · p. 19

5.4 Architecture

• The English-centric model is an 1.1B-parameter Transformer with 18 encoder and 18 decoder layers, 1024-dimensional inputs, 8192-dimensional feedforward layers, 16 attention heads, and GELU activation.The model uses pre-normalization throughout.Source: §5.4 · p. 19

5.5 Training

• Training uses an expanded FLORES-200 development set for all 22 languages and 25 language-script combinations, with validation BLEU for checkpointing and human-generated-data fine-tuning after full-corpus training.Auxiliary training generates monolingual augmentation through back-translation; downstream training uses original and back-translated parallel corpora with tagged synthetic data and the same two-stage strategy.Source: §5.5 · pp. 19–20

5.6 Data Augmentation

• Back-translated data is approximately 1.75 times the size of the original training data and is generated from English and Indic monolingual sources to diversify training.English candidates are benchmark- and training-data-deduplicated, sampled to about 400 million sentences, and allocated across language-script combinations using normalized chrF++ scores; scarce Indic-language data is included in full, while high-resource data is sampled uniformly.Source: §5.6 · pp. 20–22

5.7 Postprocessing

• Rule-based postprocessing maps script-unified En-Indic outputs back to each target language’s native Indic script; Indic-En outputs need no script conversion.Source: §5.7 · p. 22

6.1 Models Compared

• The comparison includes open and commercial systems, including a zero-shot GPT-3.5-turbo evaluation on IN22; proprietary-model results are approximations because training details are limited.Predictions from Azure, Google Translate, and GPT-3.5 were generated through their APIs using data retrieved on 10 May 2023.Source: §6.1 · p. 22

6.2 Benchmarks

• IN22 and FLORES-200 are the primary benchmarks; IN22 covers 22 languages with Wikipedia, Web, and conversational subsets, with the first two merged as IN22-Gen for evaluation.FLORES-200 includes 19 Indic languages and English-original sentences; additional benchmark results are presented in Appendix B.Source: §6.2 · p. 23

6.3 Metrics

• chrF++ is the primary metric because it better suits morphologically rich Indian languages, while COMET lacks coverage for all 22 and BLEU has tokenization and lexical-match limitations.The study also reports BLEU and COMET, and performs paired bootstrap resampling-based significance tests for all metrics.Source: §6.3 · p. 24

6.4 Generation

• IndicTrans2 generation uses beam search with beam size 5, followed by post-processing to map script-unified output to native script; open-source baselines also use beam size 5.Source sentences are preprocessed and tokenized before model input, while other baseline systems follow their documented inference procedures.Source: §6.4 · p. 24

6.5 Evaluation

• Evaluation applies standard metric processing and sacreBLEU, using Moses tokenization for English and IndicNLP or Urduhack tokenization for Indic languages; the procedure and scripts are released for reproducibility.The same evaluation procedure is used for all systems in the comparison.Source: §6.5 · p. 25

7.1 Comparison with Existing Systems

• IndicTrans2 improves over IndicTrans1 by 5.2 En-Indic and 6.3 Indic-En points on IN22-Gen, and by 5.8 and 7.4 points on FLORES-200, respectively.On IN22-Gen, it also exceeds the best commercial and open-source models for En-Indic translation by 1.7 and 4.1 points; for Indic-En, it is comparable to existing models. On FLORES-200, it is the best open-source model and performs competitively with commercial models.Source: §7.1 · pp. 25–26, Table 13 · p. 26
• IndicTrans2 consistently outperformed other open-source models across multi-domain benchmarks, except that it lagged IndicTrans1 and NLLB 1.2B on Tamil in the UFAL En-Indic test.The results also report strong gains for low-resource languages on IN22-Gen and IN22-Conv; IndicTrans2 supports 22 Indic languages and 25 language-script combinations, while performance remains weaker for low-resource Sino-Tibetan and Austroasiatic languages than for mid- and high-resource Indo-Aryan and Dravidian languages.Source: §7.1 · pp. 26–28

7.2 Understanding Data Scale vs Quality tradeoff

• Filtering mined bitext with a 0.80 LaBSE threshold reduced training data from 384M to 94M pairs, and the filtered-data model performed on par with or better than the unfiltered model.The authors adopt the filter for final training, reporting improved model performance and faster convergence after removing noisy or suboptimal pairs.Source: §7.2 · p. 29

7.3 Impact of Sequential Training with Human Annotated Data

• Seed-data fine-tuning improved average chrF++ by 2.1 En-Indic and 1 Indic-En points, but Sindhi (Arabic) En-Indic performance fell by about 3 points.The seed data includes human-annotated BPCC-H-Wiki and NLLB Seed; the authors identify it as a primary source for several mid- and low-resource languages, including Dogri, Konkani, Sindhi (Devanagari), Santali, and Manipuri (Meitei).Source: §7.3 · p. 30, Table 16 · p. 30

7.4 Impact of Data Augmentation

• Back-translation substantially improved Indic-En performance, whereas Indic back-translation did not improve En-Indic performance, leading the authors to use forward-translated data for their final En-Indic model.The authors attribute the different outcomes in part to English target-side data nearly tripling for Indic-En, while per-language target-side augmentation and domain diversity were lower for En-Indic.Source: §7.4 · pp. 31–32, Figure 6 · p. 32

7.5.1 Pivoting

• Pivot-based IndicTrans2 outperformed or matched NLLB 54B MoE across FLORES-200 Indic-to-Indic directions and significantly outperformed NLLB 1.2B distilled on IN22.The IN22 comparison uses NLLB 1.2B rather than 54B due to resource constraints; the authors treat the 1.2B results as a lower-bound estimate and anticipate a smaller gap with NLLB 54B.Source: §7.5.1 · pp. 33, 35, Table 19 · p. 33, Table 20 · p. 34

7.5.2 Direct Models

• Direct IndicTrans2-M2M uses 55.4M seed and synthetic bitext pairs across 462 directions, scoring about 1 chrF++ point below pivoting at half its inference cost.The model covers all 22 scheduled languages but lacks direct support for Kashmiri (Devanagari), Manipuri (Bengali), and Sindhi (Arabic); the compressed model has about 350M parameters and performs competitively with the 1.2B IndicTrans2-M2M.Source: §7.5.2 · p. 35

7.6 Distilled Models

• IndicTrans2-Dist models averaged drops of 0.87 chrF++ on Indic-En and 0.17 on En-Indic versus the full models across three benchmarks.The distilled models have 211M parameters versus 1.1B for IndicTrans2; Dogri, Bodo, Kashmiri, and Urdu had Indic-En drops exceeding 2.5 points, while Santali gained 2.7 points on IN22-Gen and 2.8 on IN22-Conv.Source: §7.6 · p. 35

• The authors created the first-of-their-kind multi-domain IN22 benchmark and BPCC parallel corpus, which contains about 230M bitext pairs covering 22 Indic languages.Source: §8 · p. 36
• They trained English-centric models with 1.1B parameters and compact 211M-parameter versions, plus direct Indic-Indic models with 1.2B and compact 350M parameters.The compact versions target compute-scarce settings, while the larger versions can be used in compute-heavy settings.Source: §8 · p. 36
• Their models were comparable to, if not better than, publicly available open and commercial systems across evaluations using BLEU, chrF++ (primary), and COMET.Source: §8 · p. 36
• The authors will publicly release the data, benchmarks, and model artifacts, aiming to support further advances in Indic and other low-resource language translation.Source: §8 · p. 36

Data

• Scarce human-annotated data and limited web content constrain low-resource language generalization, while Indic sentence embeddings need calibration and coverage across all 22 languages.The paper names Manipuri (Meitei), Santali, and Sindhi (Devnagari) as languages with limited generalization, and proposes calibrating embeddings with human annotations to improve mining.Source: §9 · p. 37

Modeling

• A performance gap remains between low- and high-resource languages; future work should improve cross-lingual transfer, enable efficient domain adaptation, and extend coverage without catastrophic forgetting.The models prioritize general-purpose use; the paper calls for sparse parameter-efficient domain adaptation that preserves this utility and techniques to incorporate additional Indic languages.Source: §9 · p. 37

Benchmark

• English-original n-way test sets make Indic-to-English evaluation rely on translationese; Indic-original benchmarks are needed for fair assessment (Zhang & Toral, 2019).The authors are working toward and intend to release Indic-original-to-English benchmarks for all 22 scheduled Indic languages.Source: §9 · p. 37

Evaluation

• The preliminary human evaluation covers 100 IN22-Gen sentences for English-to-Indic; broader free-form and task-oriented studies and Indic-suited automatic metrics remain future needs.The paper also calls for calibrating model-based metrics, which may not support some languages, to improve their correlation with human judgments.Source: §9 · p. 37

Fairness

• Web-collected training data may introduce social biases, and distilled models can propagate or amplify teacher-model biases (Gupta et al., 2022).The authors call for comprehensive bias identification and alignment methods to mitigate these biases and support broader, safer accessibility.Source: §9 · p. 38

Parallel Corpus Collection and Mining; Human Translation

• Parallel corpus collection and mining involved Raghavan AK, Jay Gala, and Aswanth Kumar; human translation involved seven listed contributors.The human translation contributors were Pranjal A. Chitale, Jay Gala, Mitesh M. Khapra, Pratyush Kumar, Anoop Kunchukuttan, Janki Nawale, and Anupama Sujatha.Source: §10 · p. 38

Model Training; Distillation

• Pranjal A. Chitale, Raj Dabre, Jay Gala, and Varun Gumma contributed to both model training and distillation.The same four authors are listed for both areas.Source: §10 · p. 38

Model Evaluation

• Model evaluation involved Pranjal A. Chitale, Raj Dabre, Sumanth Doddapaneni, Jay Gala, Varun Gumma, Anoop Kunchukuttan, and Ratish Puduppully.Source: §10 · p. 38

Research Leads

• Research leads were Raj Dabre, Mitesh M. Khapra, Pratyush Kumar, and Anoop Kunchukuttan.Source: §10 · p. 38

Project Conceptualization and Direction

• Project conceptualization and direction involved Mitesh M. Khapra, Pratyush Kumar, Anoop Kunchukuttan, and Vivek Raghavan.Source: §10 · p. 38

The Team Behind the Scenes

• The work acknowledges its contributors and lists administrative and Shoonya teams spanning operations, project coordination, and software development.The Shoonya roster includes project management, frontend and backend development, and technical leadership roles.Source: The Team Behind the Scenes · p. 39, Administrative Team · p. 40, Shoonya Team · p. 40

Translation Team

• The translation-team roster names language leads, senior translators, project managers, and translators across the listed language teams.The roster includes teams for languages such as Assamese, Bengali, Bodo, Dogri, Gujarati, and Hindi.Source: Translation Team · p. 41

A Data Contribution and Coverage

• Figures 7 and 8 summarize training-data contributions across seed, mined, and backtranslation axes and domain coverage for IN22-Gen and IN22-Conv.The figure caption distinguishes Indic BT, with a synthetic English side and original Indic side, from English BT, with the reverse arrangement.Source: Figure 7 · p. 57, Figure 8 · p. 57

B.1 Zero-Shot Translation Capabilities of IndicTrans2 Through Cross-Lingual Transfer

• Maithili-tag decoding performs best across Awadhi, Bhojpuri, Chhattisgarhi, and Magahi; IndicTrans2 differs from NLLB 54B MoE by around 4 points on average except for Chhattisgarhi.The comparison concerns Indic-En zero-shot evaluation on FLORES-200; Hindi and Nepali tags rank behind Maithili.Source: §B.1 · p. 58, Table 25 · p. 58

B.2 Translation Capabilities of Zero-Shot Prompted LLMs

• IndicTrans2 exceeds GPT3.5 by an average of 12.2 points in Indic-En and 14.5 points in En-Indic on the IN22 benchmark.The reported advantage is lower for high-resource Hindi (+6.6 Indic-En; +5.4 En-Indic) than for low-resource Assamese (+19.1 Indic-En; +20.4 En-Indic).Source: §B.2 · pp. 58–59, Table 26 · p. 59

B.3 Comparison with SeamlessM4T Multimodal Translation Model

• The paper reports preliminary SeamlessM4T-Large and Large v2 results on FLORES-200, IN22-Gen, and IN22-Conv.Both variants are 2.3B-parameter models, described as the best models released in that work.Source: §B.3 · p. 59

B.4 Results on NTREX

• NTREX is a news-domain benchmark covering 128 languages, including 13 scheduled Indic languages, with detailed results reported in Tables 29–31.NTREX expands the WMT 2019 test-data language coverage.Source: §B.4 · p. 59

B.5 Results on WAT2020 & WAT2021

• WAT covers translation involving 8 Indic languages in news and Hindi–English in IT and WikiNews; WAT 2021 benchmarks translation between 10 Indic languages and English.Detailed results are reported in Tables 32–37.Source: §B.5 · p. 59
• FLORES200 and IN22-Gen/IN22-Conv chrF++ tables compare SM4T, SM4Tv2, and IT2 in En-Indic and Indic-En directions.The captions define the average as covering all supported languages.Source: Table 27 · p. 60, Table 28 · p. 60

B.6 Results on WMT & UFAL

• WMT shared tasks cover Hindi (2014), Gujarati (2019), and Tamil (2020), while UFAL is an English–Tamil benchmark with cinema, news, and biblical-source sentences.Detailed results are reported in Tables 38–40.Source: §B.6 · pp. 60–61

B.8 BLEU Scores for IN22 & FLORES

• The appendix also reports chrF++, COMET, and BLEU comparisons on NTREX in En-Indic and Indic-En directions.The captions describe best-system and significant-difference markings; the supplied text does not include score values.Source: Table 29 · p. 61, Table 30 · p. 61, Table 31 · p. 62

B.7 COMET Scores for IN22 & FLORES

• COMET scores for IN22-Gen, IN22-Conv, and FLORES-200 are reported for En-Indic and Indic-En directions.The table captions mark the best-performing system and significant differences where IT2 outperforms another system.Source: §B.7 · p. 61, Table 41 · p. 65, Table 42 · p. 66

B.8 BLEU Scores for IN22 & FLORES

• BLEU scores for IN22-Gen, IN22-Conv, and the FLORES-200 devtest set are reported in En-Indic and Indic-En directions.The captions note that † marks completely off-target translations.Source: §B.8 · p. 61, Table 44 · p. 67, Table 45 · p. 68
• The WAT score tables compare systems on WAT-2020 and WAT-2021 using chrF++, COMET, and BLEU in En-Indic and Indic-En directions.The captions indicate the best system and, where applicable, significant differences favoring IT2; they do not provide the score values here.Source: Table 32 · p. 62, Table 33 · p. 63, Table 34 · p. 63

C Human Evaluation

• The small-scale blinded evaluation compared IndicTrans2, NLLB, Google, and Azure on 100 En-Indic IN22-Gen pairs across 16 languages, with results warranting caution.The sample comprised 50 Wikipedia and 50 Web examples; evaluators were fluent in English and the relevant native language, and the paper calls for broader coverage.Source: C Human Evaluation · p. 70
• Evaluators used reference-free XSTS adequacy ratings on a 1–5 scale, with calibration based on 50 English sentence pairs sampled evenly across five scoring classes.Moderated calibration adjustment was used to address extreme calibration shifts and improve cross-language score comparability.Source: C Human Evaluation · p. 70
• In selected IN22 languages, IndicTrans2 significantly outperformed other models in low-resource Konkani, Sanskrit, and Nepali; most supported languages approached a 4 XSTS rating.Hindi, Bengali, and Telugu ratings skewed high, while Bodo ratings were more symmetrical.Source: Figure 9 · p. 71, C Human Evaluation · p. 71
• IndicTrans2 ranked top, comparable with Azure and significantly better than Google and NLLB (N54B) in calibrated En-XX XSTS evaluation.The comparison uses post-calibration results and ANOVA with post-hoc Tukey HSD testing (p ≤0.05).Source: Table 47 · p. 72, C Human Evaluation · p. 72
• XSTS correlated with BLEU at 0.49 and chrF++ at 0.12 across languages, rising to 0.67 and 0.25 without Urdu; it showed no correlation with COMET.The authors partly attribute Urdu’s effect to higher fertility with UrduHack tokenization, which inflated BLEU and chrF++ scores.Source: C Human Evaluation · pp. 72–73
• IndicTrans2 XSTS scores changed by an average of 0.23 after calibration, while model rankings stayed unchanged; Punjabi and Sanskrit scores dropped, and Malayalam, Tamil, and Urdu scores improved by more than 0.2 points.Calibration also lowered Assamese scores relative to Bengali despite comparable pre-calibration scores.Source: C Human Evaluation · p. 72, Table 48 · p. 73

D Distilled Models

• Distilled students use a base18L architecture with shared decoder embeddings and output projections, yielding equal-sized directional models and saving about 60M parameters in En-Indic.The section also presents student and teacher comparisons on FLORES200 and IN22-Conv.Source: D Distilled Models · p. 74, Table 49 · p. 74, Table 50 · p. 74
• Word-level knowledge-distilled Indic-En and En-Indic models are compared using chrF++ on FLORES-200 and IN22-Conv, including seed-data fine-tuning differences versus IndicTrans2.The supplied passages describe the comparison tables but do not provide their cell values.Source: Table 52 · p. 75, Table 53 · p. 76

E Additional details about IN22 Benchmark

• IN22’s additional benchmark details cover source and domain diversity, sentence length distributions, FLORES-200 domain diversity, and conversational-subset statistics.The supplied captions identify Figure 10 as showing domain versus length distributions for Web Sources and Wikipedia, and Tables 54–55 as covering domain diversity and conversation statistics; they do not provide the underlying values.Source: E Additional details about IN22… · p. 77, Figure 10 · p. 77, Table 54 · p. 77

E.1 Source Selection

• IN22 selects Wikipedia sentences across categories with length and overlap filters, uses Indian-topic government and digital-library PDFs for Web Sources, and translates contextualized conversations into 22 Indic languages.Conversation turns are separate benchmark instances, and speaker information is metadata rather than text to translate.Source: §E.1 · pp. 78–79, Table 56 · p. 79

F.1 Translation Stages

• Translation creation proceeds through source selection, source verification, translation, and quality review, with reviewers correcting or rejecting translations for another round.The selected material spans domains including news, tourism, business, entertainment, history, geography, culture, sports, and health; experienced reviewers check each translation against guidelines.Source: §F.1 · pp. 80–81

F.2 Translation Guidelines.

• Guidelines seek natural, consistent translations across 22 languages while accommodating differences in syntax, scripts, dialects, and sociocultural features.For test sets, translators work from scratch without MT output to avoid system-output bias.Source: §F.2 · p. 81

F.3 Shoonya Translation Interface

• Shoonya supports translation and review with cross-lingual assistance for low-resource languages and a two-level maker-checker process in which experienced reviewers verify translations.Its listed capabilities also include transliteration, paragraph context, and post-editing support.Source: §F.3 · pp. 81–82

G Language of India

• The language overview uses 2011 census speaker counts and the language classification of Joshi et al. (2020), whose taxonomy ranges from extremely low-resource to high-resource.The table caption also notes that many languages are spoken across multiple states.Source: G Language of India · p. 83, Table 57 · p. 83

H Examples of language translation

• An illustrative table presents the same UN Declaration on Human Rights sentence translated into all 22 Indian languages.The source sentence is “All human beings are born free and equal in dignity and rights.”Source: Table 59 · p. 85, H Examples of language translation · p. 85

I Language Coverage of various MT models

• A table compares coverage of India’s 22 scheduled languages across open-source and commercial NMT systems.The supplied table caption gives the comparison’s scope but not system-level coverage values.Source: Table 61 · p. 87, I Language Coverage of various MT models · p. 87

J Model Card - IndicTrans2

• The authors provide an IndicTrans2 model card following Mitchell et al. (2019).Source: J Model Card - IndicTrans2 · p. 88

J.1 Model Details

• IndicTrans2 comprises two 18-layer encoder-decoder transformer models with 1.1B parameters, one for each English–Indic translation direction.The models were developed by AI4Bharat; the paper describes version 1.0.0, released May 26, 2023.Source: §J.1 · p. 88
• The models follow the paper’s training configuration and data, preprocessing, tokenization, and postprocessing steps, and are available under a permissive MIT license.The repository provides further information and accepts questions or comments through issue submissions.Source: §J.1 · p. 88

J.2 Intended Use

• IndicTrans2 supports single or batch translation across 22 Indic languages and 25 language-script combinations to and from English, plus pivot-based Indic-to-Indic translation.Intended for research and commercial use, it serves researchers, individuals translating day-to-day content, and organizations translating proprietary or internal content.Source: §J.2 · p. 88
• The models are released under the MIT license without usage limitations.Source: §J.2 · p. 88

J.3 Data, Metrics, Limitations, and Recommendations

• Training combines parallel corpora with synthetic data, while fine-tuning uses BPCC-H-Wiki and NLLB-Seed gold-standard bitext.Synthetic data generated from intermediate IndicTrans2 models augments training of the final models.Source: §J.3 · p. 89
• Evaluation covers FLORES-200 and IN22 using chrF++, BLEU, and COMET, with chrF++ as the primary metric and XSTS human evaluation on part of IN22 Combined.The baselines follow the same evaluation procedure as IndicTrans2.Source: §J.3 · p. 89
• The paper describes known caveats of the IndicTrans2 models in Section 9.The supplied passage points to that section but does not specify the caveats.Source: §J.3 · p. 89
• The authors recommend expanding support to additional languages or improving existing-language performance through minimal fine-tuning.They note that some languages beyond the currently supported 22 are not supported.Source: §J.3 · p. 89

K.1 Model Details

• Following Mitchell et al. (2019), the model card describes AI4Bharat’s IndicTrans2-M2M and Dist-M2M models as 18-layer encoder-decoder transformers for Indic-to-Indic translation.M2M has 1.2B parameters and the compact Dist-M2M variant has 350M; the models are version 1.0.0 and were released on Dec 01, 2023.Source: §K.1 · p. 90, K Model Card - IndicTrans2-M2M · p. 90
• The models use the training configuration and data described in Section 7.5.2, with preprocessing, tokenization, and postprocessing detailed in Sections 5.2, 5.3, and 5.7.IndicTrans2-M2M and Dist-M2M are made available under a permissive MIT license.Source: §K.1 · p. 90

K.2 Intended Use

• The models support single or batch translation between 22 Indic languages for research and commercial purposes, serving researchers, individual users, and organizations.The described users include researchers advancing Indic language technologies, people translating day-to-day content, and organizations translating proprietary or internal content.Source: §K.2 · p. 90
• The models are released under the MIT license without usage limitations.Source: §K.2 · p. 90

K.3 Data, Metrics, Limitations, and Recommendations

• The models use the training corpora described in Section 7.5.2 and are evaluated on benchmarks including FLORES-200 and IN22, with chrF++ as the primary metric alongside BLEU and COMET.The baselines follow the same evaluation procedure as the IndicTrans2-M2M and IndicTrans2-Dist-M2M models.Source: §K.3 · p. 91
• The authors note known model caveats and omit XSTS human evaluation for IndicTrans2-Dist, while identifying support for additional languages and minimal fine-tuning as future directions.They describe the models as foundational translation models supporting all 22 scheduled Indic languages.Source: §K.3 · p. 91

L Model Card - IndicTrans2-Dist

• Following Mitchell et al. (2019), the authors provide a model card for IndicTrans2-Dist.Source: L Model Card - IndicTrans2-Dist · p. 92

L.1 Model Details

• IndicTrans2-Dist version 1.0.0 was released on Dec 01, 2023, as 18-layer encoder-decoder transformers with 211M parameters, one for each translation direction between English and Indic.Source: §L.1 · p. 92
• The models are available under a permissive MIT license.Their training configuration, data, and preprocessing, tokenization, and postprocessing are documented in the paper’s referenced sections and appendix.Source: §L.1 · p. 92

L.2 Intended Use

• IndicTrans2-Dist is designed for research and commercial translation of text across 22 Indic languages, covering 25 language-script combinations to and from English, with pivot-based translation between Indic languages.Source: §L.2 · p. 92

L.3 Data, Metrics, Limitations, and Recommendations

• Training of IndicTrans2 models uses parallel corpora described in Section 7.6.Source: §L.3 · p. 93
• IndicTrans2-Dist fine-tuning uses BPCC-H-Wiki and NLLB-Seed, with gold-standard bitext statistics reported in Table 1.Source: §L.3 · p. 93
• Evaluation uses FLORES-200 and IN22 benchmarks, with baselines following the same generation and evaluation procedure as IndicTrans2-Dist.The procedures are described in Sections 6.4 and 6.5.Source: §L.3 · p. 93
• IndicTrans2-Dist is evaluated with chrF++, BLEU, and COMET, with chrF++ designated as the primary metric.Source: §L.3 · p. 93
• The models have known caveats, and the work does not conduct an XSTS human evaluation for IndicTrans2-Dist.The known caveats are described in Section 9.Source: §L.3 · p. 93
• Future work could expand IndicTrans2-Dist beyond its 22 supported scheduled languages or improve existing supported languages through minimal fine-tuning.The paper characterizes the model as a compact yet strong foundational translation model with extensive support for all 22 scheduled Indic languages.Source: §L.3 · p. 93

M Dataset Card

• Following Gebru et al. (2021); Pushkarna et al. (2022), the authors provide a dataset card for BPCC, used to train IndicTrans2, and IN22, the Indic-language benchmark test sets.Source: M Dataset Card · p. 94

M.1 Dataset Description

• BPCC contains approximately 230 million bitext pairs across 22 scheduled Indic languages, including mined and human-created components.BPCC-Mined has approximately 228 million pairs, including around 126 million newly mined pairs; BPCC-Human has 2.2 million gold-standard English–Indic pairs, including 644K Wikipedia pairs and 139K day-to-day-use sentences. Coverage spans 25 language-script combinations.Source: §M.1 · p. 94
• IN22 is an n-way benchmark whose IN22-Gen evaluation set combines Wikipedia and Web content, while IN22-Conv covers conversational translation.The Wikipedia and Web subsets include news, entertainment, culture, legal, and India-centric topics; IN22-Conv targets day-to-day conversational applications.Source: §M.1 · p. 94

M.2 Dataset Creation

• BPCC data draws on IndicCorp v2, the Internet Archive, Wikipedia, and in-house day-to-day-use material; IN22-Web uses government PDFs and open-source books, while IN22-Conv is annotator-created.Annotations were performed on the Shoonya platform.Source: §M.2 · p. 95
• The authors acknowledge that large-scale mining of public websites may unintentionally include personal or sensitive information.They invite concerns about potential leakages to be reported for assistance and resolution.Source: §M.2 · p. 95

M.3 Considerations for Using the Data

• The dataset is intended to improve translation quality for all 22 scheduled Indic languages, including low-resourced languages that use multiple scripts.The authors do not explicitly examine data bias in this work and identify further bias investigation as future work.Source: §M.3 · p. 95

M.4 Additional Information

• Existing mined corpora and seed corpora, as well as newly added mined corpora, are released under CC0.The named existing mined sources are NLLB and Samanantar; the existing seed sources are NLLB, ILCI, and MASSIVE, and the newly added mined corpora are Samanantar++ and Comparable.Source: §M.4 · p. 95
• Newly added Wiki and Daily seed corpora are released under CC BY, and the IN22 test set is released under CC BY 4.0.Source: §M.4 · p. 95