SearchGen-20K
收藏资源简介:
SearchGen-20K 是一个用于训练、分析和审计搜索增强图像生成行为的数据集,与题为《Search Beyond What Can Be Taught》的研究论文(arXiv:2607.05382)相关联。该研究旨在推动代理视觉生成中的知识边界,其核心机制是在图像生成前进行信息检索,以弥补生成模型的知识盲区。数据集包含 20,188 个训练提示的元数据,以及与之配套的完整生成过程追踪记录、生成的图像和统一的评估侧文件。具体规模包括:96,848 条追踪记录、283,493 张已发布的生成图像和同等数量(283,493)的评估侧文件。数据内容主要分为几个部分:1) 训练提示元数据(`train_metadata.jsonl`);2) 详细的生成过程追踪输出(`trace_artifacts.jsonl`);3) 生成的图像清单(`generation_artifacts.jsonl`,图像数据以 TAR 分片形式存储);4) 评估者输出的侧文件清单(`judge_artifacts.jsonl`);5) 用于连接生成图像与评估文件的索引清单(`generated_image_eval_inventory.tsv`)。训练提示元数据通过 `reference_slots` 等字段链接到独立的十亿级搜索语料库数据集(SearchGen-Corpus-1M),体现了“先检索后生成”的流程。内部各文件通过 `row_id`、`trace_id`、`generation_id` 等字段相互关联。该数据集主要用于搜索增强图像生成模型的训练、行为分析和算法审计。评估基准数据集(SearchGen-Bench)是独立的。
SearchGen-20K is a dataset for training, analyzing, and auditing search-augmented image generation behavior, associated with the research paper titled Search Beyond What Can Be Taught (arXiv:2607.05382). This research aims to advance the knowledge boundaries in agent-based visual generation, with a core mechanism of information retrieval before image generation to address the knowledge gaps in generative models. The dataset includes metadata for 20,188 training prompts, along with complete generation process trace records, generated images, and unified evaluation side files. The specific scale comprises: 96,848 trace records, 283,493 published generated images, and an equal number (283,493) of evaluation side files. The data content is mainly divided into several parts: 1) training prompt metadata (`train_metadata.jsonl`); 2) detailed generation process trace outputs (`trace_artifacts.jsonl`); 3) generated image inventory (`generation_artifacts.jsonl`, with image data stored in TAR shards); 4) evaluator output side file inventory (`judge_artifacts.jsonl`); 5) index inventory for linking generated images with evaluation files (`generated_image_eval_inventory.tsv`). The training prompt metadata is linked to an independent billion-scale search corpus dataset (SearchGen-Corpus-1M) through fields like `reference_slots`, reflecting the retrieve-then-generate workflow. Internal files are interconnected via fields such as `row_id`, `trace_id`, and `generation_id`. This dataset is primarily used for training search-augmented image generation models, behavior analysis, and algorithm auditing. The evaluation benchmark dataset (SearchGen-Bench) is independent.
SearchGen-20K 数据集详情
基本信息
- 任务类别: 文本到图像(Text-to-Image)
- 许可证: Apache-2.0
- 项目名称: Search Beyond What Can Be Taught: Evolving the Knowledge Boundary in Agentic Visual Generation
- 相关链接:
- arXiv: https://arxiv.org/abs/2607.05382
- 项目页面: https://haozheh3.github.io/SearchGen/
- GitHub: https://github.com/HaozheH3/SearchGen
- 关联数据集: SearchGen-Corpus-1M (https://huggingface.co/datasets/JasperHaozhe/SearchGen-Corpus-1M) 和 SearchGen-Bench (https://huggingface.co/datasets/JasperHaozhe/SearchGen-Bench)
数据集统计
| 统计项 | 数量 |
|---|---|
| 训练提示词(Prompts) | 20,188 |
| 追踪记录(Traces) | 96,848 |
| 已发布的生成图像 | 283,493 |
| 已发布的评估侧车文件 | 283,493 |
数据集用途
用于训练、分析或审计搜索增强的图像生成行为。提供提示词元数据、指向搜索语料库的参考槽链接、追踪级流水线输出、生成图像记录以及评估侧车记录。
主要文件
| 文件路径 | 说明 |
|---|---|
metadata/train_metadata.jsonl |
每行对应一个训练提示词 |
metadata/train_metadata_manifest.json |
训练元数据的来源和行数清单 |
metadata/id_summary.json |
公开ID摘要 |
metadata/trace_index.jsonl |
紧凑型追踪可用性索引 |
traces/trace_artifacts.jsonl |
每条追踪的阶段输出 |
traces/generation_artifacts.jsonl |
生成图像清单(本次不含图像载荷) |
traces/judge_artifacts.jsonl |
评估侧车记录清单 |
traces/generated_image_eval_inventory.tsv |
TSV格式,关联生成图像与评估文件 |
交叉引用关系
-
训练行与 SearchGen-Corpus-1M 的关联字段:
reference_slots[].query_idreference_slots[].entry_idreference_slots[].asset_idreference_slots[].asset_portable_pathsearch_query_ids
-
追踪和生成文件内部关联字段:
row_idtrace_idlane_idreasoner_idrow_source_pathlane_source_pathgeneration_idjudge_id
-
基准测试: 独立于本训练集,详见 SearchGen-Bench 数据集。
路径根目录说明
row_source_path和lane_source_path:相对于原始生成工件根目录。image_path:相对于最终发布的data_release/searchgen-20k/generated_artifacts/。- 评估侧车定位字段:相对于
data_release/searchgen-20k/generated_artifacts/。 composition_source_file_paths:使用相同根目录data_release/searchgen-20k/generated_artifacts/。reference_slots[].asset_portable_path:相对于../searchgen-corpus-1m/。
数据发布说明
包含283,493张生成图像以及每张图像对应的经过清洗的统一评估协议侧车文件。在Hugging Face上以确定性纯TAR分片形式存储于 data/generated/ 目录下,成员路径与发布的 generated_images/ 布局一致。分片信息见 data/generated/shards.jsonl,包含规范路径、TAR成员、大小和SHA256校验值。
权利与下架
权利问题和下架请求请联系:jasper.whz@outlook.com
引用
bibtex @article{wang2026searchgen, title = {Search Beyond What Can Be Taught: Evolving the Knowledge Boundary in Agentic Visual Generation}, author = {Wang, Haozhe and Feng, Weijia and Yu, Jinpeng and Liu, Che and Nie, Ping and Lin, Fangzhen and Liu, Jiaming and Huang, Ruihua and Lin, Jimmy and Chen, Wenhu and Wei, Cong}, journal = {arXiv preprint arXiv:2607.05382}, year = {2026}, url = {https://arxiv.org/abs/2607.05382} }





