coild-repo/coild_benchmark_v1

COILD Benchmark v1

A test set for Indic-to-Indic machine translation from the COILD project, IIT Patna: 40 translation directions, 21 languages, 2,000 sentence pairs per direction (80,000 pairs in total). Translation is direct between Indian languages, without pivoting through English.

Layout

One directory per direction, named <source>-<target> with FLORES-200 codes, holding two parallel plain-text files, one sentence per line, aligned by line number:

asm_Beng-hin_Deva/
├── test.asm_Beng     2000 lines, source
└── test.hin_Deva     2000 lines, reference

from pathlib import Path

d = Path("asm_Beng-hin_Deva")
src = d.joinpath("test.asm_Beng").read_text(encoding="utf-8").splitlines()
ref = d.joinpath("test.hin_Deva").read_text(encoding="utf-8").splitlines()
pairs = list(zip(src, ref))      # 2000 pairs, aligned by line

Languages

code	language
asm_Beng	Assamese
ben_Beng	Bengali
brx_Deva	Bodo
doi_Deva	Dogri
gom_Deva	Konkani
guj_Gujr	Gujarati
hin_Deva	Hindi
kan_Knda	Kannada
kas_Arab	Kashmiri (Arabic)
mai_Deva	Maithili
mal_Mlym	Malayalam
mar_Deva	Marathi
mni_Mtei	Manipuri (Meetei Mayek)
npi_Deva	Nepali
ory_Orya	Odia
pan_Guru	Punjabi
sat_Olck	Santali (Ol Chiki)
snd_Deva	Sindhi (Devanagari)
tam_Taml	Tamil
tel_Telu	Telugu
urd_Arab	Urdu

Directions

The 40 directions fall into two groups that do not share sentences.

Hindi-centric (34 directions) — 17 languages paired with Hindi, both ways. All of these are n-way parallel: every direction uses the same 2,000 Hindi sentences and their translations, so scores across these directions are computed on identical content and can be compared directly.

  asm_Beng-hin_Deva ben_Beng-hin_Deva brx_Deva-hin_Deva
  doi_Deva-hin_Deva gom_Deva-hin_Deva guj_Gujr-hin_Deva
  hin_Deva-asm_Beng hin_Deva-ben_Beng hin_Deva-brx_Deva
  hin_Deva-doi_Deva hin_Deva-gom_Deva hin_Deva-guj_Gujr
  hin_Deva-kas_Arab hin_Deva-mai_Deva hin_Deva-mar_Deva
  hin_Deva-mni_Mtei hin_Deva-npi_Deva hin_Deva-ory_Orya
  hin_Deva-pan_Guru hin_Deva-sat_Olck hin_Deva-snd_Deva
  hin_Deva-tel_Telu hin_Deva-urd_Arab kas_Arab-hin_Deva
  mai_Deva-hin_Deva mar_Deva-hin_Deva mni_Mtei-hin_Deva
  npi_Deva-hin_Deva ory_Orya-hin_Deva pan_Guru-hin_Deva
  sat_Olck-hin_Deva snd_Deva-hin_Deva tel_Telu-hin_Deva
  urd_Arab-hin_Deva

Dravidian (6 directions) — Tamil paired with Telugu, Kannada and Malayalam, both ways. These share a different set of 2,000 sentences, aligned through Tamil, with no overlap with the Hindi-centric group.

  kan_Knda-tam_Taml  mal_Mlym-tam_Taml  tam_Taml-kan_Knda  tam_Taml-mal_Mlym  tam_Taml-tel_Telu  tel_Telu-tam_Taml

Two consequences worth knowing:

• Scores from the two groups are not directly comparable: the sentences differ, so a gap between, say, hin_Deva-tel_Telu and tam_Taml-tel_Telu reflects content as well as language pair.
• Kannada, Malayalam and Tamil never pair with Hindi here.
• Telugu is the only language present in both groups, and its sentences are entirely different: 1,997 unique sentences with Hindi (3 lines repeat), 2,000 with Tamil, and 0 in common. That makes Telugu a useful control: the same model measured on two independent content pools. A hin_Deva-tel_Telu vs tam_Taml-tel_Telu gap is a difference in text, not in the model's Telugu ability.

Domains

Agriculture, Climate, Education, Governance, Healthcare, Judiciary, Science & Technology, Tourism.

Notes

• Files are UTF-8 with mixed line endings: 49 of the 80 files use CRLF, the rest LF, and the two files of one language can differ this way (for example the Santali file in hin_Deva-sat_Olck is CRLF while the one in sat_Olck-hin_Deva is LF, with identical content). Python's text mode handles this transparently; tools that read bytes may leave a trailing \r on a line, so compare checksums only after normalising.
• Within the Hindi-centric group, both directions of a language were checked line-by-line and carry identical Hindi and identical target sentences.
• Scores reported on this benchmark should state the metric and tokenizer. The COILD models are scored with sacreBLEU using the flores200 tokenizer, plus chrF, chrF++ and TER, against these references as shipped.