K-SEED: Korean Vision-Language Model Evaluation Benchmark

K-SEED is a Korean adaptation of the SEED-Bench evaluation suite for vision-language models, created by NCSOFT. It was developed by translating the first 20 percent of the SEED-Bench test subset into Korean and reviewing its naturalness through human inspection. The benchmark consists of questions across 12 evaluation dimensions, such as scene understanding, instance identity, and instance attribute.

Use Cases

• Benchmarking Korean vision-language model performance based on the 12 evaluation dimensions mentioned in the description
• Evaluating scene understanding capabilities of multimodal models based on Korean text and image data
• Assessing model performance on instance identity and attribute recognition tasks in a Korean context

Strengths

• Human-reviewed translation for naturalness, as stated in the description
• Covers 12 distinct evaluation dimensions, including scene understanding and instance attribute
• Specifically adapted for Korean language evaluation, addressing a linguistic gap

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download
• Column-level documentation is absent; field semantics must be inferred after download
• Row count and file formats are unknown, which may limit suitability assessment

Provenance

Source

NCSOFT

Collection Method

Adapted from SEED-Bench via translation and human review of the first 20 percent of the test subset.

Time Range

null

Freshness

Last updated 2025-07-25 06:40:38; freshness should be verified

Geography

null