WEBVTT 1 00:00:00.000 --> 00:00:02.567 To train an agent that works in a terminal, 2 00:00:02.567 --> 00:00:06.567 we first need tasks, and tasks are built from real public files. 3 00:00:06.567 --> 00:00:09.067 This is the data page of the public index; 4 00:00:09.067 --> 00:00:13.100 each point is one source, layered by the training stage it can feed. 5 00:00:13.100 --> 00:00:19.900 Under the name, the release the page reads, release-2026-11-12.13; 6 00:00:19.900 --> 00:00:22.967 every command the page gives is pinned to it. 7 00:00:22.967 --> 00:00:30.500 A drag turns the scene. 8 00:00:30.500 --> 00:00:34.400 On the left, the training stages, each with what it can use. 9 00:00:34.400 --> 00:00:38.500 We need material to build tasks from, so we pick post-training; 10 00:00:38.500 --> 00:00:41.500 its layer comes forward, and the others step back. 11 00:00:41.500 --> 00:00:48.367 On the right, the selection, 1,764 material sets, the largest first. 12 00:00:48.367 --> 00:00:50.667 Material sets come from two places: 13 00:00:50.667 --> 00:00:54.667 works the search found upstream, or a group of rows from a corpus, 14 00:00:54.667 --> 00:00:55.900 counted in units. 15 00:00:55.900 --> 00:00:59.333 A corpus is one upstream read in bulk, counted in rows. 16 00:00:59.333 --> 00:01:03.233 A first filter, Training, allowed for all rows, 17 00:01:03.233 --> 00:01:06.133 so every file may be used for training. 18 00:01:06.133 --> 00:01:11.433 The count on the right follows, 1,391 material sets. 19 00:01:11.433 --> 00:01:12.700 A second filter, 20 00:01:12.700 --> 00:01:17.800 a search for zenodo, a repository where researchers deposit their data. 21 00:01:17.800 --> 00:01:19.967 14 material sets are left. 22 00:01:19.967 --> 00:01:24.700 A click on a source opens it, with its domain, layer, form, units, 23 00:01:24.700 --> 00:01:27.900 size, license tier and training stages. 24 00:01:27.900 --> 00:01:29.567 Its point is ringed in the scene, 25 00:01:29.567 --> 00:01:32.533 and the list below the screen has its whole card. 26 00:01:32.533 --> 00:01:37.533 The same points can be arranged by domain, 27 00:01:37.533 --> 00:01:41.567 and back by training stage. 28 00:01:41.567 --> 00:01:46.300 At the bottom, the one command that takes this selection, pinned to the release. 29 00:01:46.300 --> 00:01:49.767 The index is a list of addresses, not a copy of the files; 30 00:01:49.767 --> 00:01:52.800 the command fetches each file from where it is published. 31 00:01:52.800 --> 00:01:55.333 We set units per source to at most 3, 32 00:01:55.333 --> 00:01:58.567 so the command takes up to three units from each source. 33 00:01:58.567 --> 00:02:01.367 We copy it and run it in a clean terminal. 34 00:02:01.367 --> 00:02:04.600 We paste the command as the page gave it, and run it. 35 00:02:04.600 --> 00:02:06.533 uv fetches the tool and runs it; 36 00:02:06.533 --> 00:02:09.400 the tool reads the index at the pinned release. 37 00:02:09.400 --> 00:02:12.767 Before downloading, it says what the default profile leaves out, 38 00:02:12.767 --> 00:02:14.967 then what it takes from each source. 39 00:02:14.967 --> 00:02:17.900 In all, this run takes 18 units. 40 00:02:17.900 --> 00:02:21.867 Sped up: the fetch ran for 93 seconds, shown in 20 seconds. 41 00:02:21.867 --> 00:02:25.200 Each file is checked against the size and hash the index records; 42 00:02:25.200 --> 00:02:26.554 only a match counts. 43 00:02:27.467 --> 00:02:31.767 The files are in data, a folder for each domain, and beside them _meta. 44 00:02:31.767 --> 00:02:34.867 _meta keeps the selection and each file's size and hash, 45 00:02:34.867 --> 00:02:36.967 every try and each license. 46 00:02:36.967 --> 00:02:38.500 The index gives you material; 47 00:02:38.500 --> 00:02:41.000 building tasks from it is up to you. 48 00:02:41.000 --> 00:02:45.933 From one need to checked files on disk, with the command the page gave.