Datasets:

---

XLearning-SCU

/

ARK-Bench

Like 2

Hi ARK-Bench team,

Thank you for releasing this excellent benchmark! We are evaluating several multimodal embedding models (GME-Qwen2-VL-7B, Ovis-VL-Embedding-9B, Qwen3-VL-Embedding) on ARK-Bench and would like to compute the metrics locally for research purposes.

However, we noticed that the HuggingFace dataset (XLearning-SCU/ARK-Bench) has drop_labels: true in the config, and there is no qrels file shipped with the dataset. We tried to infer the qrels by assuming (topic, query_id) == (topic, corpus_id) → positive, but this turns out to be incorrect because the query and gallery images are different files (e.g., Spatial-3D_000000.png for query vs Spatial-3D_000131.png for corpus_id=0).

Could you please clarify:

1. How is the ground-truth relevance (qrels) constructed? Specifically, for each query, which gallery item(s) are the positive candidates?
2. Is there a way to obtain the qrels for local evaluation? We understand the leaderboard uses hidden labels to prevent overfitting, but for research purposes, we would appreciate access to the ground truth (perhaps with a registration / NDA / usage agreement).
3. If local evaluation is not possible, what is the recommended workflow? Should we generate top-k predictions and submit to the leaderboard for scoring?

Context

• We are evaluating on all 17 subtypes (Biology, Physics, Chemistry, ..., Spatial-3D, Spatial-Cogmap).
• Our local setup follows the paper's protocol: per-subtype retrieval, cosine similarity, Recall@1/5/10/20, nDCG@k, MAP@k.
• Example of the mismatch we found:
  • Query 0 (Spatial-3D): query_image_file_name = images/Spatial-3D/Spatial-3D_000000.png, text = "Given the arrow indicating the viewing direction on the 3D figure above, find the 2D projection image..."
  • Gallery corpus_id=0 (Spatial-3D): file_name = images/Spatial-3D/Spatial-3D_000131.png (a different image)

Since the query and gallery images are clearly different, our inferred qrels (query_id == corpus_id) must be wrong. Any guidance on how to correctly construct or obtain the qrels would be greatly appreciated!

Thank you for your time and for maintaining this benchmark.