gemmabooru
收藏资源简介:
GemmaBooru 是一个正在进行中的、个人/公共的衍生图像-描述存档数据集,旨在构建一个便于训练的图像-描述对集合。该数据集源自在 Hugging Face 上的 `deepghs/danbooru2024` 数据集,包含了从 Danbooru 社区精选的帖子。每个数据样本包含一个全尺寸的源图像,并配有两个文本文件:一个是由 Gemma 大语言模型(通过 LM Studio 使用特定提示词生成)的自然语言描述文件(`.txt`),另一个是包含原始 Danbooru 标签的侧文件(`_booru.txt`)。此外,部分样本可能包含用于元数据/调试的 `.json` 文件。数据以分片(shard)形式组织,存储为 `.tar` 文件,并通过 Parquet 格式的元数据文件进行管理。一个全局的 `metadata.parquet` 文件提供了跨所有分片的可搜索索引,包含的关键字段有:Danbooru 帖子 ID (`post_id`)、内容分级 (`rating`, 通常为 s/q/e)、社区评分 (`score`)、收藏数 (`fav_count`)、图像尺寸 (`width`, `height`)、文件扩展名 (`file_ext`)、原始标签文本 (`booru_tags`)、Gemma 生成的描述 (`caption`) 以及文件在分片内的路径信息。数据集规模较小(标签显示少于1万个样本),主要面向研究人员和开发者,适用于图像描述生成、图像生成模型训练、描述实验以及从标签到自然语言描述的转换工作流。数据集提供了专门的 Python 工具(如 `download_from_hf.py`),支持用户根据 Danbooru 标签、内容分级、社区评分等元数据灵活地过滤和下载所需子集。重要注意事项:1. 数据集内容可能包含成人、露骨或不适合所有受众的材料,用户需自行负责内容过滤。2. 描述由模型生成,可能存在遗漏或事实错误(幻觉)。3. 原始 Danbooru 标签可能存在噪声和不一致性。4. 这是一个衍生数据集,用户在使用或再分发时应遵守上游数据源(Danbooru)的相关政策与许可要求,并注明出处。
GemmaBooru is an ongoing personal and public derivative image-caption archive dataset designed to construct a readily trainable image-caption pair collection. This dataset originates from the `deepghs/danbooru2024` dataset hosted on Hugging Face, and includes curated posts from the Danbooru community. Each data sample contains a full-size source image, paired with two text files: one is a natural language description file (`.txt`) generated by the Gemma Large Language Model (using specific prompts via LM Studio), and the other is a sidecar file (`_booru.txt`) containing the original Danbooru tags. Additionally, some samples may include `.json` files for metadata or debugging purposes. The data is organized into shards, stored as `.tar` files, and managed via Parquet-format metadata files. A global `metadata.parquet` file provides a searchable index across all shards, with key fields including: Danbooru post ID (`post_id`), content rating (`rating`, typically s/q/e), community score (`score`), favorite count (`fav_count`), image dimensions (`width`, `height`), file extension (`file_ext`), original tag text (`booru_tags`), Gemma-generated caption (`caption`), and the file path within its corresponding shard. The dataset is relatively small (fewer than 10,000 samples as indicated by its tags), primarily targeted at researchers and developers, and applicable to workflows such as image caption generation, image generation model training, captioning experiments, and tag-to-natural-language-description conversion. Specialized Python tools (e.g., `download_from_hf.py`) are provided, allowing users to flexibly filter and download desired subsets based on Danbooru tags, content ratings, community scores, and other metadata. Important notes: 1. The dataset may contain adult, explicit, or content unsuitable for all audiences, and users are solely responsible for their own content filtering. 2. The captions are generated by the model, and may contain omissions or factual errors (hallucinations). 3. The original Danbooru tags may contain noise and inconsistencies. 4. This is a derivative dataset; users must comply with the relevant policies and licensing requirements of the upstream data source (Danbooru) when using or redistributing it, and properly cite the original source.




