|
Download README.md from CheneyWong/wtw-table-annotations: direct link, hf CLI and curl.
- Browser
- Download file 1.24 kB
-
https://huggingface.co/datasets/CheneyWong/wtw-table-annotations/resolve/main/README.md
- Command line
-
hf download hf://datasets/CheneyWong/wtw-table-annotations/README.md
-
curl -L -o README.md https://huggingface.co/datasets/CheneyWong/wtw-table-annotations/resolve/main/README.md
1.24 kB
metadata
license: other
task_categories:
- table-recognition
tags:
- wtw
- table-recognition
- real-scene
- scanned-documents
size_categories: 1K<n<10K
wtw-table-annotations
WTW (Wild Table/Wild scanned documents) 表格的双标注版本: 每条样本同时给出
| column | 说明 |
|---|---|
| image | 图片字节 (struct {bytes, path}) |
| html_opus | 用 VLM(Opus) 从图片蒸馏出来的表格 HTML —— 这是训练用的标注 |
| html_wtw_original | WTW 数据集自带的官方 HTML 标注 —— 这是参照 |
背景: 2026-10-20 爬取数据时, WTW 的裁图经 VLM 蒸馏成了 HTML。2026-10-20 本机那份
中间数据被一个外部进程删掉, 图和蒸馏标注一度全丢; 后来从
WTW_image_html_pairs/*.tar 里按 id 把原图和 WTW 官方 HTML 一起找回来了
(src/recover_crawled_images.py / WTW tar 扫描), 蒸馏 HTML 则一直保存在 HF 的
CheneyWong/real-scene-tables 的 data/table_html-* 里。
两版 HTML 放在一起, 是为了能直接看「VLM 蒸馏」相对官方标注差在哪 (单元格划分、跨行跨列、空 cell 的处理) —— 这类差异在评测口径上影响很大。
同批数据的整页图版见 CheneyWong/real-scene-tables。