Datasets:
Download media/usage-example.zh.vtt from Battam/training-data-foundation: direct link, hf CLI and curl.
- Browser
- Download file 16.4 kB
-
https://huggingface.co/datasets/Battam/training-data-foundation/resolve/refs%2Fpr%2F21/media/usage-example.zh.vtt
- Command line
-
hf download hf://datasets/Battam/training-data-foundation@refs/pr/21/media/usage-example.zh.vtt
-
curl -L -o usage-example.zh.vtt https://huggingface.co/datasets/Battam/training-data-foundation/resolve/refs%2Fpr%2F21/media/usage-example.zh.vtt
16.4 kB
| WEBVTT | |
| 1 | |
| 00:00:01.150 --> 00:00:05.550 | |
| 要训练一个在终端里干活的 agent,也就是智能体, | |
| 2 | |
| 00:00:05.550 --> 00:00:07.010 | |
| 我们首先得有题。 | |
| 3 | |
| 00:00:07.210 --> 00:00:11.450 | |
| 那么一道题能不能拿来训练,可以归结成两件事, | |
| 4 | |
| 00:00:11.450 --> 00:00:16.410 | |
| 第一,环境里的材料是真实的,第二,判分靠得住, | |
| 5 | |
| 00:00:16.410 --> 00:00:20.510 | |
| 做对了才给分,什么都不做、走捷径都不给分。 | |
| 6 | |
| 00:00:20.710 --> 00:00:25.350 | |
| 这支视频我们就完整走一遍,从一个公开的数据页出发, | |
| 7 | |
| 00:00:25.350 --> 00:00:29.350 | |
| 挑材料,取文件,造出两道题,做一遍检查, | |
| 8 | |
| 00:00:29.350 --> 00:00:31.930 | |
| 最后真跑,让判分来验收。 | |
| 9 | |
| 00:00:32.130 --> 00:00:35.750 | |
| 屏幕上的命令和输出,都是当时真跑录下来的。 | |
| 10 | |
| 00:00:36.450 --> 00:00:38.630 | |
| 我们先打开数据页。 | |
| 11 | |
| 00:00:38.830 --> 00:00:41.730 | |
| 可以看到,这份索引分成两层。 | |
| 12 | |
| 00:00:41.930 --> 00:00:47.370 | |
| 上面这一层是任务材料,一共 1793 个材料集, | |
| 13 | |
| 00:00:47.370 --> 00:00:51.130 | |
| 一个材料集,就是从同一个来源挑出来的一批文件, | |
| 14 | |
| 00:00:51.130 --> 00:00:53.870 | |
| 再按任务的类别分到 13 个合集里。 | |
| 15 | |
| 00:00:54.070 --> 00:00:57.990 | |
| 下面这一层是语料库,一共 124 个, | |
| 16 | |
| 00:00:57.990 --> 00:01:01.270 | |
| 一个语料库,就是把一个上游整个读下来, | |
| 17 | |
| 00:01:01.270 --> 00:01:05.370 | |
| 加起来有 12 亿多行,在没有合适的材料集时拿来搜。 | |
| 18 | |
| 00:01:05.570 --> 00:01:08.450 | |
| 这两层各算各的,不能加在一起, | |
| 19 | |
| 00:01:08.450 --> 00:01:11.350 | |
| 材料集也不是从语料库里挑出来的。 | |
| 20 | |
| 00:01:11.550 --> 00:01:13.490 | |
| 造题要用的是上面这一层。 | |
| 21 | |
| 00:01:13.690 --> 00:01:17.930 | |
| 材料集里的文件分成单元,一个单元可以是一个工作簿, | |
| 22 | |
| 00:01:17.930 --> 00:01:22.010 | |
| 也可以是一个测试用例,一共 4171 个单元, | |
| 23 | |
| 00:01:22.010 --> 00:01:24.110 | |
| 其中 787 个带答案。 | |
| 24 | |
| 00:01:24.310 --> 00:01:28.890 | |
| 有答案才判得了分,所以这次我们就从带答案的单元里挑。 | |
| 25 | |
| 00:01:29.617 --> 00:01:31.317 | |
| 我们往下看各个合集。 | |
| 26 | |
| 00:01:31.517 --> 00:01:35.277 | |
| 带答案的电子表格这个合集,有 68 个材料集, | |
| 27 | |
| 00:01:35.277 --> 00:01:38.497 | |
| 530 个带答案的单元,我们点开它。 | |
| 28 | |
| 00:01:38.697 --> 00:01:42.477 | |
| 我们搜一下英国政府发布的工作簿,把这一行展开。 | |
| 29 | |
| 00:01:45.995 --> 00:01:49.995 | |
| 它收的是英国政府机构发布的表格,一共有 40 个单元, | |
| 30 | |
| 00:01:49.995 --> 00:01:52.075 | |
| 一个单元就是一个工作簿, | |
| 31 | |
| 00:01:52.075 --> 00:01:54.895 | |
| 40 个答案单独放在 answer 目录下面。 | |
| 32 | |
| 00:01:55.095 --> 00:01:59.255 | |
| 许可证是 OGL UK 3.0,属于宽松的这一类, | |
| 33 | |
| 00:01:59.255 --> 00:02:00.955 | |
| 训练和分发都允许。 | |
| 34 | |
| 00:02:01.155 --> 00:02:02.455 | |
| 再点开预览, | |
| 35 | |
| 00:02:04.522 --> 00:02:09.642 | |
| 可以看到这份索引只是一份清单,不是文件的副本, | |
| 36 | |
| 00:02:09.642 --> 00:02:12.702 | |
| 每个文件一行,记着从哪下载、 | |
| 37 | |
| 00:02:12.902 --> 00:02:13.802 | |
| 多大、 | |
| 38 | |
| 00:02:14.002 --> 00:02:15.142 | |
| 校验值 | |
| 39 | |
| 00:02:15.342 --> 00:02:16.402 | |
| 和许可证。 | |
| 40 | |
| 00:02:18.117 --> 00:02:22.137 | |
| 再用同样的办法,从磁盘镜像这个合集里挑一个。 | |
| 41 | |
| 00:02:22.337 --> 00:02:26.197 | |
| 这是 exfatprogs 这个工具自带的测试用例, | |
| 42 | |
| 00:02:26.521 --> 00:02:30.841 | |
| 一共 15 个损坏的 exFAT 镜像,每个都带答案, | |
| 43 | |
| 00:02:30.841 --> 00:02:34.761 | |
| 镜像本身的许可证是 GPL,属于相同方式共享, | |
| 44 | |
| 00:02:34.761 --> 00:02:39.101 | |
| 训练允许,分发这些镜像时要带上同样的许可证。 | |
| 45 | |
| 00:02:39.817 --> 00:02:42.957 | |
| 两个都看好了,我们来到获取数据这一节。 | |
| 46 | |
| 00:02:43.157 --> 00:02:48.217 | |
| 页面一次只给当前这一个材料集拼命令,先预览,再获取。 | |
| 47 | |
| 00:02:48.417 --> 00:02:52.257 | |
| 命令里钉死了一个发布版本,清单和校验值就定了, | |
| 48 | |
| 00:02:52.257 --> 00:02:56.657 | |
| 以后再取,要么取到一模一样的文件,要么校验不过, | |
| 49 | |
| 00:02:56.657 --> 00:02:57.797 | |
| 直接报错。 | |
| 50 | |
| 00:02:57.997 --> 00:02:59.917 | |
| 我们要的是两个材料集, | |
| 51 | |
| 00:02:59.917 --> 00:03:02.897 | |
| 下面就在终端里把它们写进同一条命令。 | |
| 52 | |
| 00:03:03.817 --> 00:03:06.777 | |
| 如果是 agent 来挑,或者你不想开网页, | |
| 53 | |
| 00:03:06.777 --> 00:03:10.057 | |
| 索引里有一张总表,就是这个文件, | |
| 54 | |
| 00:03:10.057 --> 00:03:13.657 | |
| 每个材料集和语料库各占一行,用 DuckDB, | |
| 55 | |
| 00:03:13.657 --> 00:03:17.757 | |
| 一个能直接查数据文件的数据库工具,一句 SQL 就能查。 | |
| 56 | |
| 00:03:17.957 --> 00:03:22.437 | |
| 总表里合集用字母代号,带答案的电子表格是 A, | |
| 57 | |
| 00:03:22.437 --> 00:03:27.657 | |
| 磁盘镜像是 B,同样只看带答案的,一共有 111 个材料集。 | |
| 58 | |
| 00:03:27.857 --> 00:03:31.857 | |
| 每个合集我们列出单元最多的 6 个,刚才挑的这个, | |
| 59 | |
| 00:03:31.857 --> 00:03:36.117 | |
| 和这个,都在里面,许可证的等级也写在这一列。 | |
| 60 | |
| 00:03:37.637 --> 00:03:40.597 | |
| 取文件用的是索引自带的取用工具, | |
| 61 | |
| 00:03:40.597 --> 00:03:44.217 | |
| uv 会把它直接从网上拉下来运行,不用先安装。 | |
| 62 | |
| 00:03:44.417 --> 00:03:48.257 | |
| 两个来源写在同一条命令里,每个先取一个单元, | |
| 63 | |
| 00:03:48.257 --> 00:03:50.897 | |
| 再加上 dry run,也就是只预览, | |
| 64 | |
| 00:03:50.897 --> 00:03:55.317 | |
| 不下载,没选中的文件太多,这里就过滤掉不显示了。 | |
| 65 | |
| 00:03:55.517 --> 00:03:58.817 | |
| 可以看到,这两个来源一共 55 个单元。 | |
| 66 | |
| 00:03:59.017 --> 00:04:03.177 | |
| 取用工具默认会去掉重复的、和评测基准重合的、 | |
| 67 | |
| 00:04:03.177 --> 00:04:07.497 | |
| 还没和基准比对过的、许可不允许分发的、不是英文的、 | |
| 68 | |
| 00:04:07.497 --> 00:04:11.757 | |
| 带个人数据的,还有超过 20 MB 的,这次都是 0。 | |
| 69 | |
| 00:04:11.957 --> 00:04:15.557 | |
| 按我们的要求取 2 个单元,连答案和配套文件, | |
| 70 | |
| 00:04:15.557 --> 00:04:16.857 | |
| 一共 8 个文件。 | |
| 71 | |
| 00:04:17.057 --> 00:04:20.677 | |
| 确认没问题,我们去掉 dry run,真正下载。 | |
| 72 | |
| 00:04:21.388 --> 00:04:23.148 | |
| 8 个文件全部 ok, | |
| 73 | |
| 00:04:23.148 --> 00:04:26.288 | |
| 每个都和索引里记的大小和校验值核对过。 | |
| 74 | |
| 00:04:26.488 --> 00:04:30.568 | |
| 其中这个 exFAT 镜像,上游是放在压缩包里的, | |
| 75 | |
| 00:04:30.568 --> 00:04:32.428 | |
| 工具自动把它取了出来。 | |
| 76 | |
| 00:04:32.628 --> 00:04:34.968 | |
| 我们看一下取到了什么。 | |
| 77 | |
| 00:04:35.168 --> 00:04:37.328 | |
| 材料放在按领域分的目录下面, | |
| 78 | |
| 00:04:37.328 --> 00:04:39.348 | |
| 答案单独放在 answer 下面。 | |
| 79 | |
| 00:04:39.548 --> 00:04:42.988 | |
| 之所以分开放,是因为答案是用来判分的, | |
| 80 | |
| 00:04:42.988 --> 00:04:44.528 | |
| 不能进 agent 的环境。 | |
| 81 | |
| 00:04:45.250 --> 00:04:49.970 | |
| 索引交给大家的是材料和答案,造题是使用者自己的事, | |
| 82 | |
| 00:04:49.970 --> 00:04:54.450 | |
| 下面演示的是其中一种做法,按 Harbor 的任务格式来造, | |
| 83 | |
| 00:04:54.450 --> 00:04:56.870 | |
| Harbor 是专门跑这类任务的框架。 | |
| 84 | |
| 00:04:57.070 --> 00:05:02.290 | |
| 一道题就是四样东西,题面、环境、参考解和判分。 | |
| 85 | |
| 00:05:04.170 --> 00:05:06.990 | |
| 那么我们先来造表格这道题。 | |
| 86 | |
| 00:05:07.190 --> 00:05:11.030 | |
| 这个 py 命令和这两个路径都写在用例的说明里, | |
| 87 | |
| 00:05:11.030 --> 00:05:14.010 | |
| 路径就是刚才取到的答案文件和工作簿。 | |
| 88 | |
| 00:05:14.210 --> 00:05:15.590 | |
| 打开答案文件, | |
| 89 | |
| 00:05:16.467 --> 00:05:18.867 | |
| 里面记的是 5 个公式格子的值, | |
| 90 | |
| 00:05:18.867 --> 00:05:23.367 | |
| 6.5、6.6、6.55、4.9、5.8。 | |
| 91 | |
| 00:05:23.567 --> 00:05:25.027 | |
| 再看工作簿本身, | |
| 92 | |
| 00:05:26.120 --> 00:05:30.140 | |
| 这 5 个格子的公式旁边,都存着上一次算出来的结果。 | |
| 93 | |
| 00:05:32.140 --> 00:05:33.820 | |
| 和答案一模一样, | |
| 94 | |
| 00:05:33.820 --> 00:05:37.900 | |
| 因为答案文件本来就是从这些存着的结果里取出来的, | |
| 95 | |
| 00:05:37.900 --> 00:05:40.560 | |
| 它们是发布时算好存进去的值。 | |
| 96 | |
| 00:05:40.760 --> 00:05:44.040 | |
| 也就是说,答案其实就藏在工作簿里, | |
| 97 | |
| 00:05:44.040 --> 00:05:47.800 | |
| 如果把这个文件原样交给 agent,它打开就能读到, | |
| 98 | |
| 00:05:47.800 --> 00:05:49.100 | |
| 这道题就白出了。 | |
| 99 | |
| 00:05:50.017 --> 00:05:51.777 | |
| 所以给 agent 的那一份, | |
| 100 | |
| 00:05:51.777 --> 00:05:55.397 | |
| 我们用 openpyxl 打开再保存一次,就这一行。 | |
| 101 | |
| 00:05:55.597 --> 00:05:59.197 | |
| openpyxl 是读写 Excel 文件的 Python 库, | |
| 102 | |
| 00:05:59.197 --> 00:06:02.337 | |
| 它保存时只写公式,不写算好的结果。 | |
| 103 | |
| 00:06:05.753 --> 00:06:06.893 | |
| 再看一遍, | |
| 104 | |
| 00:06:08.737 --> 00:06:13.157 | |
| 公式都还在,存着的结果全变成了 None,也就是没有了。 | |
| 105 | |
| 00:06:13.357 --> 00:06:17.777 | |
| 答案文件我们放进 tests 目录,只有判分的时候才用。 | |
| 106 | |
| 00:06:18.683 --> 00:06:21.963 | |
| 题面就这么几句,请算出这 5 个格子, | |
| 107 | |
| 00:06:21.963 --> 00:06:24.463 | |
| 写进 values.json 这个文件。 | |
| 108 | |
| 00:06:24.663 --> 00:06:27.783 | |
| 参考解,也就是这道题的标准做法, | |
| 109 | |
| 00:06:27.783 --> 00:06:31.783 | |
| 是用 LibreOffice 在没有界面的模式下打开工作簿, | |
| 110 | |
| 00:06:31.783 --> 00:06:35.003 | |
| 另存一份,再读出这 5 个数。 | |
| 111 | |
| 00:06:35.203 --> 00:06:40.823 | |
| 存着的结果已经去掉了,所以它打开时会把这些格子算出来。 | |
| 112 | |
| 00:06:41.750 --> 00:06:45.270 | |
| 判分就很直接了,拿答案文件里的值逐个比, | |
| 113 | |
| 00:06:45.270 --> 00:06:48.330 | |
| 相对误差在百万分之一以内才算对。 | |
| 114 | |
| 00:06:51.983 --> 00:06:57.423 | |
| 并且判分是在另一个容器里跑的,容器就是隔开的运行环境, | |
| 115 | |
| 00:06:57.423 --> 00:06:59.583 | |
| 答案只在判分的容器里, | |
| 116 | |
| 00:06:59.583 --> 00:07:02.963 | |
| 两边之间只传 values.json 这一个文件。 | |
| 117 | |
| 00:07:03.163 --> 00:07:04.843 | |
| 两边也都断了网, | |
| 118 | |
| 00:07:04.843 --> 00:07:08.763 | |
| 不然 agent 可以直接去上游把原来的文件下回来, | |
| 119 | |
| 00:07:08.763 --> 00:07:10.223 | |
| 读出存着的结果。 | |
| 120 | |
| 00:07:11.150 --> 00:07:13.090 | |
| 第二道题的答案是另一种。 | |
| 121 | |
| 00:07:13.290 --> 00:07:17.130 | |
| 我们先用 fsck,也就是文件系统的检查工具, | |
| 122 | |
| 00:07:17.130 --> 00:07:18.510 | |
| 看一下这个镜像。 | |
| 123 | |
| 00:07:18.710 --> 00:07:22.470 | |
| 可以看到报了一连串错误,文件的簇链断了, | |
| 124 | |
| 00:07:22.470 --> 00:07:26.630 | |
| 也就是文件在盘上的块接不起来了,校验和也不对, | |
| 125 | |
| 00:07:26.630 --> 00:07:29.130 | |
| 退出码是 4,确实是坏的。 | |
| 126 | |
| 00:07:29.330 --> 00:07:33.170 | |
| 它的答案文件里不是一个值,而是一条出处, | |
| 127 | |
| 00:07:33.170 --> 00:07:36.210 | |
| 写明了上游测试脚本的地址和行号, | |
| 128 | |
| 00:07:36.210 --> 00:07:41.030 | |
| 也就是 exfatprogs 自己是怎么判断修好了没有的。 | |
| 129 | |
| 00:07:41.230 --> 00:07:46.190 | |
| 镜像放进环境给 agent 修,同时在 tests 里留一份原图, | |
| 130 | |
| 00:07:46.190 --> 00:07:49.170 | |
| 判分时用它确认这道题本身是坏的。 | |
| 131 | |
| 00:07:50.083 --> 00:07:54.083 | |
| 所以题面就是,把它修到 fsck 查不出问题, | |
| 132 | |
| 00:07:54.083 --> 00:07:56.983 | |
| 修好的写到这个路径,大小不能变。 | |
| 133 | |
| 00:07:57.183 --> 00:07:58.563 | |
| 判分写了四条, | |
| 134 | |
| 00:07:59.112 --> 00:08:04.872 | |
| 第一,原来的镜像确实报错,第二,修好的镜像大小没变, | |
| 135 | |
| 00:08:04.872 --> 00:08:06.872 | |
| 第三,再查一遍, | |
| 136 | |
| 00:08:06.872 --> 00:08:11.212 | |
| 退出码是 0,第一和第三照的是上游的断言。 | |
| 137 | |
| 00:08:11.412 --> 00:08:13.012 | |
| 第四是我们加的, | |
| 138 | |
| 00:08:13.012 --> 00:08:17.012 | |
| 名字以 l0 开头的三个文件和一个目录必须还在, | |
| 139 | |
| 00:08:17.012 --> 00:08:21.492 | |
| fsck 没有报它们有错,不然重新格式化一个空镜像, | |
| 140 | |
| 00:08:21.492 --> 00:08:23.032 | |
| 前三条也都能过。 | |
| 141 | |
| 00:08:23.950 --> 00:08:28.910 | |
| 参考解也很短,用 fsck 加上修复的选项修一遍, | |
| 142 | |
| 00:08:28.910 --> 00:08:31.650 | |
| 再查一遍,然后放到要求的位置。 | |
| 143 | |
| 00:08:33.350 --> 00:08:34.810 | |
| 两道题就造好了。 | |
| 144 | |
| 00:08:35.010 --> 00:08:40.290 | |
| 我们看一下目录,每道题都有题面、环境、参考解和判分, | |
| 145 | |
| 00:08:40.290 --> 00:08:42.150 | |
| 这就是 Harbor 的任务格式。 | |
| 146 | |
| 00:08:42.350 --> 00:08:44.750 | |
| 我们还照 Terminal-Bench 4.0 的规矩, | |
| 147 | |
| 00:08:44.750 --> 00:08:49.170 | |
| 给每道题配了一份说明,和一个作弊的解,后面会用到。 | |
| 148 | |
| 00:08:51.789 --> 00:08:55.629 | |
| 我们用的版本是钉死的,先用它做一遍静态检查, | |
| 149 | |
| 00:08:55.629 --> 00:08:59.869 | |
| 也就是不运行题目,只查目录和配置合不合规矩, | |
| 150 | |
| 00:08:59.869 --> 00:09:04.189 | |
| 这里多开了两条规则,agent 不能用 root 身份跑, | |
| 151 | |
| 00:09:04.189 --> 00:09:07.149 | |
| 环境用的容器镜像也要按摘要钉死, | |
| 152 | |
| 00:09:07.149 --> 00:09:09.489 | |
| 也就是写明它的 sha256。 | |
| 153 | |
| 00:09:09.689 --> 00:09:12.509 | |
| 没有报任何问题,退出码是 0。 | |
| 154 | |
| 00:09:13.217 --> 00:09:15.237 | |
| 接下来就是真跑了。 | |
| 155 | |
| 00:09:15.437 --> 00:09:17.197 | |
| 先让 oracle 来做, | |
| 156 | |
| 00:09:17.197 --> 00:09:20.097 | |
| oracle 就是直接执行参考解的 agent。 | |
| 157 | |
| 00:09:20.297 --> 00:09:22.457 | |
| Harbor 会为每道题构建环境, | |
| 158 | |
| 00:09:22.457 --> 00:09:25.597 | |
| 跑参考解,再到单独的容器里判分。 | |
| 159 | |
| 00:09:25.797 --> 00:09:29.577 | |
| 这一段我们快进,实际用了 1 分 46 秒。 | |
| 160 | |
| 00:09:32.077 --> 00:09:34.577 | |
| 两道题平均 1 分,都通过了。 | |
| 161 | |
| 00:09:36.717 --> 00:09:38.257 | |
| 可是光这样还不够。 | |
| 162 | |
| 00:09:38.457 --> 00:09:42.137 | |
| 如果判分本身是松的,什么都不做也能拿分, | |
| 163 | |
| 00:09:42.137 --> 00:09:44.637 | |
| 那参考解过了也说明不了什么。 | |
| 164 | |
| 00:09:44.837 --> 00:09:48.777 | |
| 所以我们再让 nop 来做,nop 就是什么都不做的 agent。 | |
| 165 | |
| 00:09:51.277 --> 00:09:53.857 | |
| 两道题都是 0 分,这就对了。 | |
| 166 | |
| 00:09:54.583 --> 00:10:00.123 | |
| 什么都不做拿不到分还不够,还要看走捷径能不能骗过判分。 | |
| 167 | |
| 00:10:00.323 --> 00:10:02.643 | |
| 每道题都带了一个作弊的解, | |
| 168 | |
| 00:10:02.643 --> 00:10:06.403 | |
| exFAT 这道题直接格式化一个新的镜像, | |
| 169 | |
| 00:10:06.403 --> 00:10:10.903 | |
| 表格这道题去网上下载原来的表,抄里面存着的结果。 | |
| 170 | |
| 00:10:11.103 --> 00:10:15.103 | |
| 照用例里的说明,把作弊的解放到参考解的位置, | |
| 171 | |
| 00:10:15.103 --> 00:10:16.643 | |
| 再让 oracle 跑一遍。 | |
| 172 | |
| 00:10:20.684 --> 00:10:22.304 | |
| 两道题也都是 0 分。 | |
| 173 | |
| 00:10:24.017 --> 00:10:28.017 | |
| 我们看一下每道题的记录,参考解那一轮都是 1, | |
| 174 | |
| 00:10:28.017 --> 00:10:29.797 | |
| 后面两轮都是 0。 | |
| 175 | |
| 00:10:29.997 --> 00:10:31.297 | |
| 表格这道题, | |
| 176 | |
| 00:10:32.077 --> 00:10:35.117 | |
| 算出来的 5 个数和答案完全一致, | |
| 177 | |
| 00:10:35.117 --> 00:10:37.997 | |
| nop 和作弊这两轮都没有结果文件, | |
| 178 | |
| 00:10:37.997 --> 00:10:42.977 | |
| 直接判错,作弊的解要联网下载,可题目是断网的。 | |
| 179 | |
| 00:10:43.177 --> 00:10:46.237 | |
| exFAT 这道题, | |
| 180 | |
| 00:10:46.437 --> 00:10:50.597 | |
| 参考解四条全过,nop 那一轮后三条过不了, | |
| 181 | |
| 00:10:50.597 --> 00:10:54.757 | |
| 作弊的新镜像过得了前三条,过不了第四条, | |
| 182 | |
| 00:10:54.757 --> 00:10:56.697 | |
| 这就是我们加第四条的原因。 | |
| 183 | |
| 00:11:02.550 --> 00:11:06.390 | |
| 到这里,我们就从公开的数据页走到了两道题, | |
| 184 | |
| 00:11:06.390 --> 00:11:10.150 | |
| 参考解能过,什么都不做和作弊都过不了, | |
| 185 | |
| 00:11:10.150 --> 00:11:12.890 | |
| 这是一道题能拿去训练的起码要求。 | |
| 186 | |
| 00:11:13.090 --> 00:11:18.690 | |
| 回头看一遍,挑材料,取文件,把答案从环境里拿掉, | |
| 187 | |
| 00:11:18.690 --> 00:11:21.650 | |
| 写好题面和判分,做静态检查, | |
| 188 | |
| 00:11:21.650 --> 00:11:25.350 | |
| 最后 oracle、nop 和作弊各跑一遍。 | |
| 189 | |
| 00:11:25.550 --> 00:11:28.450 | |
| 换成别的材料集,也是这一套流程。 | |
| 190 | |
| 00:11:28.650 --> 00:11:31.530 | |
| 数据页和用例的地址就在屏幕上, | |
| 191 | |
| 00:11:31.530 --> 00:11:35.870 | |
| 这两道题和全部命令都在用例里,大家可以照着试一试。 | |