Datasets:
Download media/usage-example.zh.srt from Battam/training-data-foundation: direct link, hf CLI and curl.
- Browser
- Download file 3.54 kB
-
https://huggingface.co/datasets/Battam/training-data-foundation/resolve/main/media/usage-example.zh.srt
- Command line
-
hf download hf://datasets/Battam/training-data-foundation/media/usage-example.zh.srt
-
curl -L -o usage-example.zh.srt https://huggingface.co/datasets/Battam/training-data-foundation/resolve/main/media/usage-example.zh.srt
3.54 kB
| 1 | |
| 00:00:00,000 --> 00:00:06,633 | |
| 要训练一个在终端里干活的智能体, | |
| 首先得有题,而题要用真实的公开文件来造。 | |
| 2 | |
| 00:00:06,633 --> 00:00:12,667 | |
| 这是公开索引的数据页,每个光点是一个来源, | |
| 按它能用于的训练阶段分层。 | |
| 3 | |
| 00:00:12,667 --> 00:00:18,200 | |
| 名字下面是本页读的发布 | |
| release-2026-11-12.13, | |
| 4 | |
| 00:00:18,200 --> 00:00:20,933 | |
| 页面给出的每条命令都钉在这一版上。 | |
| 5 | |
| 00:00:20,933 --> 00:00:27,967 | |
| 拖一下,场景就转起来。 | |
| 6 | |
| 00:00:27,967 --> 00:00:31,433 | |
| 左边是三个训练阶段,各写着能用多少。 | |
| 7 | |
| 00:00:31,433 --> 00:00:37,867 | |
| 我们要的是造题的材料,所以点后训练, | |
| 它这一层亮到前面,其他几层退到后面。 | |
| 8 | |
| 00:00:37,867 --> 00:00:42,967 | |
| 右边是当前选择, | |
| 1,764 个材料集,最大的排在前面。 | |
| 9 | |
| 00:00:42,967 --> 00:00:44,467 | |
| 材料集有两个来处: | |
| 10 | |
| 00:00:44,467 --> 00:00:49,333 | |
| 检索在上游找到的作品,或一组语料库行, | |
| 按单元计。 | |
| 11 | |
| 00:00:49,333 --> 00:00:52,700 | |
| 语料库是整批读的完整上游,按行计。 | |
| 12 | |
| 00:00:52,700 --> 00:00:58,833 | |
| 第一个筛选,训练选「允许,所有行」, | |
| 每个文件都许可拿去训练。 | |
| 13 | |
| 00:00:58,833 --> 00:01:02,867 | |
| 右边的数字跟着变,剩 1,391 个材料集。 | |
| 14 | |
| 00:01:02,867 --> 00:01:08,867 | |
| 第二个筛选是检索,输入 zenodo, | |
| 一个研究者存放数据的仓库。 | |
| 15 | |
| 00:01:08,867 --> 00:01:10,267 | |
| 剩 14 个。 | |
| 16 | |
| 00:01:10,267 --> 00:01:18,833 | |
| 点一个来源就打开它,有领域、所在层、 | |
| 形式、单元数、大小、许可等级和训练阶段。 | |
| 17 | |
| 00:01:18,833 --> 00:01:23,567 | |
| 场景里它的光点被圈出来, | |
| 屏幕下方的列表里有它的完整卡片。 | |
| 18 | |
| 00:01:23,567 --> 00:01:28,133 | |
| 同一批光点可以按领域重排, | |
| 19 | |
| 00:01:28,133 --> 00:01:32,133 | |
| 再按训练阶段排回去。 | |
| 20 | |
| 00:01:32,133 --> 00:01:36,400 | |
| 底部是取回这批选择的那条命令, | |
| 钉在这一版发布上。 | |
| 21 | |
| 00:01:36,400 --> 00:01:42,600 | |
| 索引只是一份地址清单,不是文件的副本, | |
| 命令从每个文件发布的地方取回它。 | |
| 22 | |
| 00:01:42,600 --> 00:01:47,733 | |
| 「每个来源最多取」选 3, | |
| 命令就从每个来源最多取三个单元。 | |
| 23 | |
| 00:01:47,733 --> 00:01:51,000 | |
| 复制它,到一个干净的终端里跑。 | |
| 24 | |
| 00:01:51,000 --> 00:01:53,900 | |
| 把页面给的命令原样贴进来,回车运行。 | |
| 25 | |
| 00:01:53,900 --> 00:01:58,900 | |
| uv 取来取件工具并运行它, | |
| 工具读的是钉住的那一版索引。 | |
| 26 | |
| 00:01:58,900 --> 00:02:01,367 | |
| 下载前它先说默认规则排除了什么, | |
| 27 | |
| 00:02:01,367 --> 00:02:03,600 | |
| 再说从每个来源取多少。 | |
| 28 | |
| 00:02:03,600 --> 00:02:05,833 | |
| 这一次一共取 18 个单元。 | |
| 29 | |
| 00:02:05,833 --> 00:02:14,533 | |
| 这一段加速了:取回实际跑了 131 秒, | |
| 这里只放 20 秒。 | |
| 30 | |
| 00:02:14,533 --> 00:02:18,799 | |
| 每个文件都按索引记的大小和哈希核对, | |
| 对上的才算取到。 | |
| 31 | |
| 00:02:19,367 --> 00:02:23,467 | |
| 文件在 data 里, | |
| 每个领域一个文件夹,旁边是 _meta。 | |
| 32 | |
| 00:02:23,467 --> 00:02:27,100 | |
| _meta 里记着这次的选择, | |
| 和每个文件的大小与哈希; | |
| 33 | |
| 00:02:27,100 --> 00:02:30,033 | |
| 也记着每次取件和各来源的许可。 | |
| 34 | |
| 00:02:30,033 --> 00:02:31,400 | |
| 索引交的是材料, | |
| 35 | |
| 00:02:31,400 --> 00:02:34,100 | |
| 拿它造题是使用者自己的事。 | |
| 36 | |
| 00:02:34,100 --> 00:02:40,033 | |
| 从一个需求到盘上核对过的文件, | |
| 用的就是页面给的这条命令。 | |