COILD: An Indic-Centric Parallel Corpus and Benchmark for Machine Translation Across Indian Languages
Kshetrimayum Boynao Singh, Nitin Kumar Mishra, Palash Pratim Dutta, Atai Waris Khan, Aparna Kaushik, Avinash Kumar, Deeksha, Deepak Kumar, et al.
cs.CL
Quick verdict
Claims 1.16M human-verified Indic-pivot MT pairs across 20 language pairs; valuable resource, but pivot-superiority claim lacks an English-pivot control.
Summary
This resource paper presents COILD, a parallel corpus of 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 families, built entirely from originally Indian-language sources rather than English-pivot mining. A dual-pivot design (Hindi for Indo-Aryan, Tibeto-Burman and Austro-Asiatic; Tamil for Dravidian), OCR extraction, translator screening, and multi-stage expert review with a 20% TER edit-rate acceptance threshold produce an eight-domain corpus. The authors also release a 2,000-sentence expert-verified benchmark with aligned sentence identifiers (Education 600; other domains 200 each). Fine-tuning IndicTrans2-Distilled and NLLB-200 on COILD reportedly improves quality across most of the 40 translation directions on five automatic metrics and human adequacy/fluency ratings. However, the automatic-metric results appear only in figures without absolute scores or significance tests; Table 3 shows notable human-evaluation regressions (e.g., IndicTrans2 Hindi→Konkani adequacy drops 1.08) that sit uneasily beside the 'consistent improvements' narrative; the claimed superiority over English-pivot data is never tested against a matched control; and Odia human evaluation remains incomplete. Overall a valuable resource whose comparative claims need stronger evidence.
Review complete
Assessment at a glance
0
risk flags identified
● high 0
● medium 0
● low 0
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- COILD provides over 1.16 million fully human-translated and human-verified sentence pairs across 20 Indian language pairs and eight real-world domains, built entirely from original Indian-language sources.
- The dual-pivot strategy (Hindi for Indo-Aryan/Tibeto-Burman/Austro-Asiatic, Tamil for Dravidian) preserves family-internal linguistic characteristics better than English-pivot construction, and per the Introduction 'substantially improves translation quality compared to conventional English-centric approaches'.
- Fine-tuning IndicTrans2-Distilled and NLLB-200 on COILD consistently improves translation quality across most language pairs, domains, automatic metrics and human evaluation.
- The 2,000-sentence expert-verified benchmark with cross-lingually aligned sentence IDs enables systematic multilingual, many-to-many and cross-lingual evaluation for Indian languages.
- Per the Introduction, the human evaluation reveals important limitations of widely used automatic metrics for Indian languages — a claim the Results section (§5.3) actually argues against.
Strengths
Fully human-translated, human-verified data
Every sentence pair is manually translated (MT prohibited, §3.3) and independently expert-reviewed (§3.5) with a quantitative batch acceptance gate (<20% TER edit rate), directly addressing the web-crawl noise documented for Samanantar/BPCC/NLLB-style resources (Kreutzer et al., 2022).
Coverage of genuinely low-resource languages
Table 1 includes Santali (20,603 pairs), Bodo (47,599), Manipuri (52,494), Konkani (49,631), Dogri (58,896) and Maithili (93,748) — languages with almost no human-verified parallel data elsewhere, plus a translator-recruitment pipeline (§3.3) adapted to their scarcity.
Aligned-ID cross-lingual benchmark design
Typologically motivated dual-pivot construction
Dual-model validation with complementary human evaluation
Weaknesses
Headline comparative claim is untested
Automatic results locked in figures
Narrative vs. Table 3 mismatch
No inferential statistics
Incomplete, selectively explained evaluation
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
COILD (arXiv:2609.28826, v1, Sep 2026) is a resource paper from a six-institution Indian consortium (IIT Patna, IIT Delhi, IIT Guwahati, IIIT Delhi, MIT-MAHE, IGDTUW) funded under the MeitY/BHASHINI COIL-D project. It contributes: (1) a parallel corpus of 1.16M+ human-translated, human-verified sentence pairs over 20 Indian language pairs — Hindi paired with 17 languages, Tamil with Kannada/Malayalam/Telugu, 40 directions total — spanning eight domains; (2) a 2,000-sentence expert-verified benchmark (Education 600, other seven domains 200 each, Fig. 2) with sentence IDs aligned across all languages; (3) evidence that fine-tuning IndicTrans2-Distilled and NLLB-200 on this data improves quality. The Introduction adds a fourth claim: typologically appropriate pivots 'substantially improve translation quality compared to conventional English-centric approaches' — a comparative claim the experiments never actually test.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 0 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the COILD dataset and how large is it?
COILD is a parallel corpus for Indian-language machine translation containing over 1.16 million human-translated and human-verified sentence pairs across 20 Indian language pairs — Hindi paired with 17 languages, and Tamil paired with Kannada, Malayalam and Telugu — totalling about 44.79 million tokens. It spans eight real-world domains including governance, healthcare, judiciary, education, agriculture, science and technology, tourism and climate.
How is COILD different from BPCC or Samanantar?
Established corpora such as Samanantar, NLLB and BPCC are largely built by mining web text and aligning sentences through English as a pivot. COILD is built from documents originally written in Indian languages: every sentence pair is translated by a screened human translator (machine translation prohibited), then reviewed by independent linguistic experts with a batch-level 20% TER edit-rate acceptance threshold, and finally validated by domain experts.
Did fine-tuning on COILD actually improve translation quality?
The paper reports gains for most of the 40 translation directions on both IndicTrans2-Distilled and NLLB-200 across five automatic metrics (BLEU, chrF++, TER, BERTScore, COMET) and human adequacy/fluency ratings. Caveats: the automatic-metric numbers appear only in figures without absolute scores or significance tests, and a few directions regressed in human evaluation — most notably Hindi→Konkani for IndicTrans2, where adequacy fell by 1.08 points on a 5-point scale.
Where can I download COILD, and does it include a test benchmark?
The corpus is on Hugging Face under 'coild-dataset/COILD-MT-Corpus' with models under 'COIL-D/models', plus a Google Forms data-request link. The benchmark consists of 2,000 expert-verified sentences whose identifiers are aligned across all 20 languages, enabling evaluation of any Indian language pair without creating new references.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 3–6 (automatic metric values, domain heatmaps, absolute baseline/finetuned scores) are not analyzable from the extracted text; the magnitudes of the claimed automatic improvements could not be verified
- Token counts in Table 1 use Indian digit notation (lakhs); extraction formatting may introduce ambiguity in individual cells
- Odia human evaluation is explicitly incomplete in the paper itself (marked 'underway')
- Review based on arXiv v1 (23 Sep 2026); no per-pair domain proportions of the training corpus are stated anywhere in the text
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.