Evaluation

We have evaluated the performance of jina-clip-v2 on a selection of benchmarks: English and multilingual cross-modal retrieval benchmarks in Section 4.1, common text embedding tasks from the MTEB benchmark suite in Section 4.2, visual document retrieval tasks in Section 4.3, and an ablation study of jina-clip-v2’s matryoshka representations in Section 4.4.

Detailed evaluation results are presented in the appendices.

Detailed CLIP Benchmark results are given in Appendix tables 7, 8 and 9, MTEB results are in in Appendix 10. ViDoRe benchmark results are in Appendix 11.

A full MRL evaluation can be found in tables 12, 13, 14, 15 in the appendices.