Datasets:
You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset re-distributes code-switching evaluation data from many sources, each under its own license or terms, which are listed per subset on this card. By requesting access you confirm that you already have access to the original datasets of the subsets you use (or are entitled to obtain them under their terms), so that receiving them from this re-distribution is covered by that access.
Log in or Sign Up to review the conditions and access this dataset content.
cs-lid-eval: code-switching language identification evaluation
Code-switched (CS) and monolingual evaluation data for language identification (LID), gathered from the resources
used in MaskLID (Kargaran et al., ACL 2024), Burchell et al. (EACL 2024), ILP4LID (Radoła et al., Findings of EMNLP
2026) and MixLID (Hong et al., 2026), in one schema with GlotLID labels (ISO 639-3 code and ISO 15924 script).
It is for evaluation only: use the dev split to choose settings and report on test.
Schema
| field | type | description |
|---|---|---|
id |
string | {subset}-{split}-{index} |
subset, split |
string | config name; dev or test |
text |
string | the sentence (for token subsets: the tokens joined by spaces) |
tokens |
list of string or null | the source's tokens (token subsets only) |
token_tags |
list of string or null | the source's original tag of each token (e.g. lang1, ne, other, mixed) |
token_langs |
list of string/null or null | GlotLID label of each token; null for tokens that belong to no language (named entities, mixed words, punctuation, symbols, "other") |
languages |
list of string | the set of languages in the sentence (GlotLID labels) |
is_cs |
bool | the sentence contains more than one language |
source |
string | the original resource and the collection it was taken from |
Tasks: (1) CS detection (is_cs), (2) language-set prediction (languages), (3) word-level LID (token_langs,
token subsets only).
Metrics used in the literature
- CS detection: precision, recall, F1 of predicting more than one language.
- Language set: exact match (MaskLID, ILP4LID, Burchell et al.), partial match and false positives per CS label (MaskLID), sample-averaged set precision / recall / F1 and Hamming loss (Burchell et al., ILP4LID), each reported separately for CS and monolingual sentences (ILP4LID).
- Word level: token accuracy and macro F1 per sentence, averaged per language pair (MixLID); precision / recall / F1 per language and the share of unassigned words (ILP4LID).
Subsets
Subsets are named {collection}_{lang1}-{lang2}. The original corpora (sagt_, itu_, barik_, kallaama_,
basco_, euskanol_, ascend_, guaspa_, fame_, melid_, denglisch_) are converted from their official
releases at pinned revisions, keeping their own dev/test splits where they have them (otherwise even/odd sentences;
BaSCo by corpus entry, so a sentence and its rewrites stay in one split). cmb_* subsets come from CodeMixBench's LID folder (Yang and Chai, EMNLP 2025), which has no split: even-numbered
sentences of each file are dev, odd-numbered test. ilp4lid_* subsets keep ILP4LID's split (which filters
sentences to 20-200 characters and removes emojis, URLs and @usernames). ilp4lid_mono holds the monolingual control
sentences (FLORES+, the 500 shortest sentences per language, Hindi and Nepali romanized; and LinCE monolingual
sentences). The cmb_* sets also contain monolingual sentences (is_cs = false).
Overlap: cmb_hin-eng, cmb_npi-eng, cmb_spa-eng, cmb_arb-arz and ilp4lid_hin-eng, ilp4lid_npi-eng,
ilp4lid_spa-eng all derive from LinCE and may share sentences; cmb_gug-spa and guaspa_gug-spa are the same GUA-SPA data,
cmb_fry-nld and fame_fry-nld the same FAME treebank, and the ilp4lid_* code-switched subsets were built from
some of the original corpora included here (Turkish-English, Turkish-German, BaSCo/EuskanolDS, Indonesian-English,
Kallaama). Do not average overlapping subsets as independent benchmarks.
GlotLID v4 has no label for Hokkien in Han script (nan_Hani, in cmb_nan-cmn).
| subset | labels | source | dev sentences / CS | test sentences / CS | license / terms of the source |
|---|---|---|---|---|---|
ascend_cmn-eng |
sentence | ASCEND (Lovenia et al., 2022) | 1130 / 213 | 1315 / 373 | CC BY-SA 4.0 (ASCEND) |
barik_ind-eng |
token + sentence | Indonesian-English code-mixed tweets (Barik et al., 2019) | 413 / 387 | 412 / 384 | research purposes only, no license stated (tweets) |
basco_eus-spa |
sentence | BaSCo corpus (Aguirre et al., 2022), valid utterances | 1140 / 676 | 1164 / 701 | CC BY-SA 3.0 ES (BaSCo) |
cmb_arb-arz |
token + sentence | LinCE (Aguilar et al., 2020) development set, via CodeMixBench | 551 / 75 | 556 / 86 | LinCE terms (not verified; tweets, X/Twitter terms apply) |
cmb_cmn-eng |
token + sentence | Calvillo et al. 2020, via CodeMixBench (Yang and Chai, 2025) | 1511 / 1495 | 1511 / 1503 | none stated (Calvillo et al. 2020) |
cmb_deu-eng |
token + sentence | TongueSwitcher test set (Sterner and Teufel, 2023), via CodeMixBench | 626 / 413 | 626 / 419 | CC BY 4.0 (TongueSwitcher; restricted access on Zenodo; tweets) |
cmb_fry-nld |
token + sentence | Braggaar and van der Goot 2021 (UD Frisian-Dutch FAME), via CodeMixBench | 125 / 125 | 125 / 125 | CC BY-SA 4.0 (UD Frisian-Dutch FAME) |
cmb_gug-spa |
token + sentence | GUA-SPA (Chiruzzo et al., 2023), via CodeMixBench | 89 / 70 | 90 / 72 | MIT (GUA-SPA) |
cmb_hin-eng |
token + sentence | LinCE (Aguilar et al., 2020) development set, via CodeMixBench | 369 / 152 | 370 / 170 | LinCE terms (not verified; tweets/Facebook, platform terms apply) |
cmb_mar-eng |
token + sentence | L3Cube-MeLID (Chavan et al., 2023), via CodeMixBench | 670 / 431 | 670 / 409 | CC BY-NC-SA 4.0, research only (L3Cube-MeLID; tweets) |
cmb_nan-cmn |
token + sentence | Lu et al. 2022 (Taiwanese Hokkien-Mandarin), via CodeMixBench | 1900 / 1739 | 1900 / 1727 | none stated (Lu et al. 2022; synthesized text) |
cmb_npi-eng |
token + sentence | LinCE (Aguilar et al., 2020) development set, via CodeMixBench | 666 / 458 | 666 / 485 | LinCE terms (not verified; tweets) |
cmb_spa-eng |
token + sentence | LinCE (Aguilar et al., 2020) development set, via CodeMixBench | 1626 / 564 | 1621 / 578 | LinCE terms (not verified; tweets) |
denglisch_deu-eng |
token + sentence | Denglisch corpus (Osmelak and Wintner, 2023), manually annotated part | 2065 / 648 | 2068 / 629 | none stated (Denglisch corpus) |
euskanol_eus-spa |
sentence | EuskanolDS gold (Heredia et al., 2025) | 464 / 464 | 463 / 463 | CC BY-NC 4.0, and the licenses of the corpora it is derived from (EuskanolDS) |
fame_fry-nld |
token + sentence | FAME treebank (Braggaar and van der Goot, 2021), UD_Frisian_Dutch-Fame | 200 / 200 | 200 / 200 | CC BY-SA 4.0 (UD_Frisian_Dutch-Fame) |
guaspa_gug-spa |
token + sentence | GUA-SPA (Chiruzzo et al., 2023), IberLEF 2023 task 1 | 180 / 135 | 179 / 142 | MIT (GUA-SPA) |
ilp4lid_deu-tur |
sentence | Turkish-German CS data (Cetinoglu 2016 / SAGT), via ILP4LID | 100 / 100 | 1143 / 1143 | SAGT: CC BY-SA 4.0 / CC BY-NC-SA 4.0 (conflicting statements; treat as non-commercial) |
ilp4lid_eus-spa |
sentence | BaSCo (Aguirre et al., 2022) and EuskanolDS (Heredia et al., 2025), via ILP4LID | 100 / 100 | 965 / 965 | BaSCo: CC BY-SA 3.0 ES; EuskanolDS: CC BY-NC 4.0 (and the licenses of its sources) |
ilp4lid_fra-wol |
sentence | Kallaama (Gauthier et al., 2024), via ILP4LID | 0 / 0 | 4386 / 4386 | CC BY 4.0 (Kallaama) |
ilp4lid_hin-eng |
sentence | LinCE (Aguilar et al., 2020), via ILP4LID | 100 / 100 | 1928 / 1928 | LinCE terms (not verified; tweets/Facebook) |
ilp4lid_ind-eng |
sentence | Barik et al. 2019, via ILP4LID | 0 / 0 | 627 / 627 | research purposes only, no license stated (Barik et al. 2019; tweets) |
ilp4lid_mono |
sentence | FLORES+ (500 shortest sentences per language; Hindi and Nepali romanized with uroman), via ILP4LID | 500 / 0 | 7174 / 0 | CC BY-SA 4.0 (FLORES+; do not re-host where crawlers can reach it) and LinCE terms for its LinCE monolingual part |
ilp4lid_npi-eng |
sentence | LinCE (Aguilar et al., 2020), via ILP4LID | 100 / 100 | 6959 / 6959 | LinCE terms (not verified; tweets) |
ilp4lid_spa-eng |
sentence | LinCE (Aguilar et al., 2020), via ILP4LID | 0 / 0 | 8269 / 8269 | LinCE terms (not verified; tweets) |
ilp4lid_tur-eng |
sentence | Yirmibesoglu and Eryigit 2018, via ILP4LID (Radola et al., 2026) | 100 / 100 | 256 / 256 | CC BY-NC-SA 4.0 (Yirmibesoglu and Eryigit 2018; official access by request form) |
itu_tur-eng |
token + sentence | Turkish-English CS corpus (Yirmibesoglu and Eryigit, 2018) | 189 / 185 | 188 / 187 | CC BY-NC-SA 4.0 (ITU NLP group; official access by request form; tweets and Ekşi Sözlük text) |
kallaama_fra-wol |
token + sentence | Kallaama Wolof transcriptions (Gauthier et al., 2024), checked set | 3614 / 2163 | 3614 / 2166 | CC BY 4.0 (Kallaama) |
melid_mar-eng |
token + sentence | L3Cube-MeLID (Chavan et al., 2023) | 1154 / 705 | 1341 / 841 | CC BY-NC-SA 4.0, research only (L3Cube-MeLID; tweets) |
sagt_tur-deu |
token + sentence | SAGT treebank (Cetinoglu and Coltekin, 2022), UD_Turkish_German-SAGT | 801 / 746 | 805 / 762 | CC BY-SA 4.0 (LICENSE.txt) / CC BY-NC-SA 4.0 (README) of UD_Turkish_German-SAGT; treat as non-commercial |
Licenses and terms
This dataset is a collection: each subset keeps the license and terms of its original source, listed above. Access is gated: requesters confirm that they already have access to the original datasets, so this re-distribution only gives them, in one schema, data they are entitled to. Some sources are non-commercial or research-only, some state no license, and several contain social-media posts whose platform terms (e.g. X/Twitter) also apply. Please cite the original resources (see their papers) and the collections they were taken from (CodeMixBench, ILP4LID). If you are an author of a source and want its subset changed or removed, open a discussion on this repository.
- Downloads last month
- 696