omneity-labs/lid-benchmark

LID Benchmark — Language Identification Evaluation Results

Structured evaluation results for 10 language identification models across 8 benchmarks covering 380 languages — with per-language accuracy, aggregate metrics, and confusion analysis.

Built as part of the Gherbal evaluation pipeline.

The full PDF report is also available.

Quick Start

from datasets import load_dataset

# Per-language results (26,540 rows)
per_lang = load_dataset("omneity-labs/lid-benchmark", "results_per_language", split="train")

# Aggregate metrics per model × benchmark × scope (400 rows)
aggregate = load_dataset("omneity-labs/lid-benchmark", "results_aggregate", split="train")

# Summary — one row per model × benchmark, full scope only (80 rows)
summary = load_dataset("omneity-labs/lid-benchmark", "results_summary", split="train")

Leaderboard (Full Scope)

Model	FLORES+ devtest	MADAR	Gherbal-Multi	ATLASIA-LID
GlotLID	0.9253	0.5648	0.7772	0.4977
OpenLID v2	0.8748	0.6262	0.7762	0.5735
OpenLID v3 (HPLT-LID)	0.8556	—	0.6619	—
Gherbal v4	0.8500	0.6298	0.8699	0.6909
OpenLID v1	0.8425	0.5587	0.8296	0.4845
NLLB-LID	0.8331	0.1052	0.7522	0.3348
FastLID-176	0.4006	0.1352	0.6472	0.3899
Gherbal v3	0.3605	0.5745	0.8966	0.6561
Gherbal v2	0.1495	0.5811	0.7961	0.6561
Gherbal v1	0.1374	0.2771	0.8385	0.2718

Note: Full-scope FLORES+ accuracy penalizes models that support fewer languages (unsupported languages count as errors). Use results_aggregate with scope=v4 (214 languages) for a fairer narrower comparison. Gherbal v4 achieves 0.9312 accuracy on FLORES+ devtest in the v4 scope.

Dataset Configs

results_per_language — Per-Language Breakdown

26,540 rows. One row per (model, benchmark, scope, language).

Column	Type	Description
model	string	Model name (e.g. gherbal-v4, glotlid)
benchmark	string	Benchmark name (e.g. flores-devtest)
scope	string	Language scope: full, v1, v2, v3, or v4
language	string	Language code in iso639-3_Script format (e.g. arb_Arab)
n_samples	int	Number of test samples for this language
accuracy	float	Classification accuracy (0–1)
top_confusion_1	string	Most confused-with language
top_confusion_1_count	int	Count of samples misclassified as this language
top_confusion_2	string	2nd most confused-with language
top_confusion_2_count	int	—
top_confusion_3	string	3rd most confused-with language
top_confusion_3_count	int	—
confusions_json	string	Full confusion map as JSON (all misclassified targets and counts)

Example — find the hardest languages for a model:

from datasets import load_dataset
import pandas as pd

ds = load_dataset("omneity-labs/lid-benchmark", "results_per_language", split="train")
df = ds.to_pandas()

# Worst-performing languages for Gherbal v4 on FLORES
worst = (
    df[(df["model"] == "gherbal-v4") &
       (df["benchmark"] == "flores-devtest") &
       (df["scope"] == "full") &
       (df["n_samples"] >= 100)]
    .sort_values("accuracy")
    .head(10)
    [["language", "accuracy", "n_samples", "top_confusion_1"]]
)
print(worst)

Example — compare Arabic dialect accuracy across models:

arabic_dialects = [
    "arz_Arab", "ary_Arab", "arq_Arab", "aeb_Arab",
    "apc_Arab", "acm_Arab", "ars_Arab", "afb_Arab", 
    # Add more
]

arabic_df = df[
    (df["language"].isin(arabic_dialects)) &
    (df["benchmark"] == "flores-devtest") &
    (df["scope"] == "full")
]

pivot = arabic_df.pivot_table(
    index="language", columns="model", values="accuracy"
)
print(pivot.round(3))

results_aggregate — Aggregate Metrics

400 rows. One row per (model, benchmark, scope).

Column	Type	Description
model	string	Model name
benchmark	string	Benchmark name
scope	string	Language scope
accuracy	float	Overall accuracy
f1_macro	float	Macro-averaged F1
f1_weighted	float	Weighted F1
precision_macro	float	Macro-averaged precision
recall_macro	float	Macro-averaged recall
n_samples	int	Total evaluation samples
n_classes	int	Number of unique languages

Example — model comparison across scopes:

ds = load_dataset("omneity-labs/lid-benchmark", "results_aggregate", split="train")
df = ds.to_pandas()

comparison = df[
    (df["benchmark"] == "flores-devtest") &
    (df["model"].isin(["gherbal-v4", "glotlid", "openlid-v2"]))
].pivot_table(index="scope", columns="model", values="accuracy")
print(comparison.round(4))

results_summary — Quick Summary

80 rows. One row per (model, benchmark) — full scope only. Best for quick leaderboard construction.

Column	Type	Description
model	string	Model name
benchmark	string	Benchmark name
accuracy	float	Overall accuracy (full scope)
f1_macro	float	Macro F1 (full scope)
f1_weighted	float	Weighted F1 (full scope)
precision_macro	float	Macro precision (full scope)
recall_macro	float	Macro recall (full scope)
n_samples	int	Total samples
n_classes	int	Number of classes

Models Evaluated

Model	Type	Languages	Source
Gherbal v4	FastText	214	Omneity Labs
Gherbal v3	FastText	106	Omneity Labs
Gherbal v2	FastText	46	Omneity Labs
Gherbal v1	FastText	36	Omneity Labs
GlotLID v3	FastText	2,102	LMU Munich
NLLB-LID	FastText	218	Meta
OpenLID v1	FastText	201	Laurie Burchell
OpenLID v2	FastText	201	Laurie Burchell
OpenLID v3 (HPLT-LID)	FastText	201	HPLT
FastLID-176	FastText	176	Meta

Benchmarks

Benchmark	Samples	Languages	Description
FLORES+ devtest	222,640	214	openlanguagedata/flores_plus devtest split
FLORES+ dev	224,325	220	openlanguagedata/flores_plus dev split
MADAR	5,600	15	sawalni-ai/madar — Arabic dialect corpus
Gherbal-Multi	185,000+	106+	sawalni-ai/gherbal-multi — multi-source test set
ATLASIA-LID	234,000+	24	atlasia/Arabic-LID-Leaderboard — Arabic country-level dialects
WiLI-2018	—	235	Wikipedia Language Identification
CommonLID	—	—	Common Crawl language ID
Bouquet	—	—	Cross-domain evaluation

Evaluation Scopes

Results include multiple scopes to enable fair comparison between models with different language coverage:

Scope	Languages	Description
full	All	All languages in the benchmark (penalizes models with fewer supported languages)
v1	36	Intersection with Gherbal v1 language set
v2	46	Intersection with Gherbal v2 language set
v3	106	Intersection with Gherbal v3 language set
v4	214	Intersection with Gherbal v4 language set

Using scoped evaluation ensures models are compared only on languages they were designed to handle. For example, Gherbal v3 supports 106 languages — its v3 scope accuracy on FLORES+ is much higher than its full scope accuracy, because the full scope includes 108+ languages it was never trained on.

Language Codes

Languages use the iso639-3_Script format from FLORES+:

• arb_Arab — Modern Standard Arabic (Arabic script)
• arz_Arab — Egyptian Arabic
• ary_Arab — Moroccan Arabic (Arabic script)
• ary_Latn — Moroccan Arabic (Latin script)
• eng_Latn — English
• fra_Latn — French

Full list of 380 languages available in the results_per_language config.

CSV Downloads

For convenience, CSV versions of all three configs are also included in the csv/ directory.

Citation

If you use this benchmark data in your research, please reference:

• Omneity Labs LID Benchmark: https://huggingface.co/datasets/omneity-labs/lid-benchmark
• Gherbal model: https://www.omneitylabs.com/models/gherbal
• Evaluation benchmarks: See individual benchmark datasets linked above.

License

The evaluation results in this dataset are released under Apache 2.0. The underlying benchmark datasets retain their original licenses.