[KongchangAI]
Benchmark

Belebele

A multilingual reading comprehension benchmark covering more than 100 languages, with questions written by native speakers. Because of its broad language coverage and human-authored questions, it is also frequently borrowed for retrieval evaluation.

Timeline (last 90 days)

Nov 15

The final model scored an average of 17.18% on the Belebele Dravidian language evaluation, with a best score of 25.00%.

Unverified50%
Nov 10

Without fine-tuning, EmbeddingGemma-300M achieves only 51.7% to 60.8% recall@1 on the Belebele dataset for Tajik, Kazakh, Uzbek and Kyrgyz.

Unverified50%
Nov 10

On the Belebele dataset, Saidzoda-CA-EmbeddingGemma-300M improved recall@1 for the four languages by 14.4 to 26.0 percentage points, reaching 75.2% to 77.7%.

Unverified50%
Nov 10

Belebele is a multilingual reading comprehension dataset covering 122 languages, with each question accompanied by four candidate passages, commonly used to evaluate the cross-lingual retrieval ability of multilingual embedding models.

Unverified50%
Nov 3

On the Belebele test set, Saidzoda-Tajik-E5-Large's top-10 retrieval accuracy improved from 92.2% for the baseline to 95.2%, but the confidence intervals overlap.

Unverified50%

All Facts (5)

Source Articles