Datasets:
eval_annot_qa_part2
Web images found through Bing image search with queries derived from the topics of public VLM evaluation benchmarks, each with a detailed caption and a multi-turn QA conversation written by an annotator VLM.
This repo continues Elliot-Data/eval_annot_qa. That repo is the snapshot taken on 2026-08-20 for ElliotVL training (1,729,382 images). This lane has 2,003,730 annotated image files in total. 2,003,718 of them are byte-identical to an image in a snapshot repo and are left out here. The other 12 files hold 12 distinct images, which are the rows of this repo (1 parquet file, 2.9 MB).
- Images whose SHA-256 appears in
wordnet_annot_qaoreval_annot_qaare left out, so this repo and the snapshots do not overlap. This holds even when such an image received a new annotation later. - No two rows share an image. When the same image (byte-identical) was found by several queries, it is one
row whose
qaholds the QA of all its annotations in order, the same rule as in the snapshot. - These rows were not part of the training snapshot.
- They are crawled web images, not benchmark images, but they are topically close to those benchmarks (
clean_meta.benchmarksnames the benchmark a query came from). Check overlap before reporting results on them. - Self-labelled by the ElliotVL annotation pipeline: a Qwen-397B vision-language model wrote the caption and QA from the image under a structured-JSON prompt that also scores image quality, richness and
query_relevance(kept per row inclean_meta). The page title and description of the source page were given to the annotator as hints. No judge, verifier or human reviewed these rows, so treat them as model-generated supervision.
Fields
image: HFImage(), the image bytes as downloadedimage_sha256: SHA-256 of the image bytesqa: list of {question,answer}, in order. A row with several turns is one conversation over the image.n_qa: number of turnsclean_meta: JSON string with the annotator's scores (image_quality,aesthetics,richness,entity_count,entity_type_count,query_relevancefrom 0 to 10, and a per-categorycategory_relevance), itsdetailed_caption, and the searchqueryand pagetitlethat found the imagefamily,canonical_source: the same values as ineval_annot_qa, so the two repos can be mixed directly
Rights
The images were collected from the public web and belong to their owners. They are shared for research
only, behind manual access approval. Do not use an image where its original licence does not allow it. To
have an image removed, open a discussion with its image_sha256.
How this was made
The concept tree, crawl queries, annotation prompt and annotation client are in elliotvl-data-pipeline (stage 6, self-labelled lanes).
- Downloads last month
- 22