wikiart-zero-shot

# Dataset Card for Wikiart Zero Shot Queries

<!-- Provide a quick summary of the dataset. -->

Wikiart Zero Shot Queries is a tiny slice of the wikiart dataset with added zero shot queries, which can be used to qualitatively test inference of zero shot object detection models such as [Grounding Dino](https://huggingface.co/docs/transformers/en/model_doc/grounding-dino).

## Dataset Creation
This dataset contains the first 1k entries of the [wikiart dataset](https://huggingface.co/datasets/huggan/wikiart), with added zero shot queries. The queries were obtained using [LLaVA-NeXT](https://huggingface.co/docs/transformers/en/model_doc/llava_next) with the vicuna-13b backbone ([weights](https://huggingface.co/llava-hf/llava-v1.6-vicuna-13b-hf)), and the following prompt:
```
"A chat between a curious human and an artificial intelligence assistant. The assistant responds to the human's instructions. USER: <image>\nList things in the image, limit yourself to the most important things only. Answer like this, no repetitions 'a cow. a dog. a hat' ASSISTANT:"
```

# Wikiart 零样本查询数据集卡片

<!-- 请提供该数据集的简要概述。 -->

Wikiart 零样本查询（Zero Shot Queries）是wikiart数据集的小型子集，新增了零样本查询内容，可用于定性测试如[Grounding Dino](https://huggingface.co/docs/transformers/en/model_doc/grounding-dino)这类零样本目标检测模型（Zero Shot Object Detection）的推理性能。

## 数据集构建
本数据集选取了[wikiart数据集](https://huggingface.co/datasets/huggan/wikiart)的前1000条条目，并新增了零样本查询。本次生成的查询内容由搭载vicuna-13b骨干网络的[LLaVA-NeXT](https://huggingface.co/docs/transformers/en/model_doc/llava_next)完成，所用权重来自[llava-v1.6-vicuna-13b-hf](https://huggingface.co/llava-hf/llava-v1.6-vicuna-13b-hf)，使用的提示词如下：

"一位好奇的人类与人工智能助手的对话。助手将遵照人类的指示进行回复。用户：<图片>
请列出图像中的物体，仅保留最核心的内容。请按照如下格式作答，请勿重复：‘一头奶牛。一只狗。一顶帽子’ 助手："