training-data-foundation / media /usage-example.en.vtt
Battam's picture
regenerate the index pages only (generator be62deea, templates 2824f53e) [run publish-20261105T071924Z]
13a5b61 verified
Raw History Blame
17.7 kB
WEBVTT
1
00:00:01.150 --> 00:00:04.690
To train an agent that works
in a terminal, we first need tasks.
2
00:00:04.890 --> 00:00:08.510
Whether a task is good enough
to train on comes down to two things.
3
00:00:08.710 --> 00:00:11.450
First, the material
in its environment is real.
4
00:00:11.650 --> 00:00:16.370
Second, the grading can be trusted:
you score only when you get it right,
5
00:00:16.370 --> 00:00:19.590
and doing nothing
or taking a shortcut scores nothing.
6
00:00:19.790 --> 00:00:22.370
This video walks through
the whole process once.
7
00:00:22.570 --> 00:00:27.610
From a public data page, we pick material,
fetch the files, build two tasks,
8
00:00:27.610 --> 00:00:31.310
check them, and run them for real,
so the grading has the last word.
9
00:00:31.510 --> 00:00:34.890
Every command and output
on screen comes from a real run.
10
00:00:35.617 --> 00:00:37.877
First, we open the data page.
11
00:00:38.077 --> 00:00:40.417
You can see the index has two layers.
12
00:00:40.617 --> 00:00:42.957
The top layer is task material.
13
00:00:43.157 --> 00:00:46.777
There are 1,793 material sets.
14
00:00:46.977 --> 00:00:50.277
A material set is a batch of files picked
from one source.
15
00:00:50.477 --> 00:00:54.177
They are grouped by task family
into 13 collections.
16
00:00:54.377 --> 00:00:57.997
The bottom layer is corpora, 124 of them.
17
00:00:58.197 --> 00:01:01.097
A corpus is a whole upstream,
read in bulk.
18
00:01:01.297 --> 00:01:06.757
Together they hold 1.22 billion rows,
for when no material set fits.
19
00:01:06.957 --> 00:01:10.577
The two layers are counted separately,
so don't add them up.
20
00:01:10.777 --> 00:01:13.517
And material sets
are not taken from the corpora.
21
00:01:13.717 --> 00:01:15.817
Tasks are built from the top layer.
22
00:01:16.017 --> 00:01:18.917
The files in a material set
are split into units.
23
00:01:19.117 --> 00:01:21.697
A unit can be a workbook, or a test case.
24
00:01:21.897 --> 00:01:27.677
There are 4,171 units,
and 787 have an answer file.
25
00:01:27.877 --> 00:01:32.617
Without an answer, you can't grade,
so we pick from units with answers.
26
00:01:33.317 --> 00:01:35.257
Let's scroll down to the collections.
27
00:01:35.457 --> 00:01:37.397
Here's Spreadsheets with answers.
28
00:01:37.597 --> 00:01:42.017
It has 68 material sets,
and 530 units with an answer file.
29
00:01:42.217 --> 00:01:43.357
Let's open it.
30
00:01:43.557 --> 00:01:47.097
We search for the UK government workbooks,
and expand that row.
31
00:01:47.297 --> 00:01:50.437
It holds spreadsheets published
by UK government bodies.
32
00:01:50.888 --> 00:01:54.428
There are 40 units,
and each unit is one workbook.
33
00:01:54.628 --> 00:01:57.608
The 40 answers are kept apart,
in the answer folder.
34
00:01:57.808 --> 00:02:03.508
The license is OGL UK 3.0,
in the Permissive tier.
35
00:02:03.708 --> 00:02:06.208
Training and redistribution
are both allowed.
36
00:02:06.408 --> 00:02:07.868
Now we open Preview.
37
00:02:09.775 --> 00:02:13.715
You can see the index is only a list,
not a copy of the files.
38
00:02:13.915 --> 00:02:18.015
Each file gets one row:
where to download it, its size,
39
00:02:18.215 --> 00:02:19.435
its checksum,
40
00:02:19.635 --> 00:02:21.095
and its license.
41
00:02:22.817 --> 00:02:26.437
In the same way, we pick
one from the Disk images collection.
42
00:02:26.637 --> 00:02:29.777
These are the test cases
that come with the exfatprogs tool.
43
00:02:30.741 --> 00:02:34.121
There are 15 damaged exFAT images,
each with an answer.
44
00:02:34.321 --> 00:02:37.861
The images themselves are under the GPL,
in the Share-alike tier.
45
00:02:38.061 --> 00:02:39.361
Training is allowed.
46
00:02:39.561 --> 00:02:43.261
If you ship the images, you have to ship
the same license with them.
47
00:02:43.983 --> 00:02:46.803
With both chosen, we go to Get the data.
48
00:02:47.003 --> 00:02:50.523
The page builds commands
for one material set at a time:
49
00:02:50.523 --> 00:02:53.183
Look first to preview, then Take to fetch.
50
00:02:53.383 --> 00:02:57.483
The commands pin a release, which fixes
the file list and the checksums.
51
00:02:57.683 --> 00:03:01.703
Fetch again later, and you get
the exact same files, or a checksum error.
52
00:03:01.903 --> 00:03:04.863
We want two material sets,
so in the terminal
53
00:03:04.863 --> 00:03:06.803
we'll put both into one command.
54
00:03:07.717 --> 00:03:11.637
If an agent does the picking,
or you'd rather not open the page,
55
00:03:11.637 --> 00:03:13.257
the index has a catalog.
56
00:03:13.457 --> 00:03:17.797
It's this file: one row
for each material set and each corpus.
57
00:03:17.997 --> 00:03:20.077
DuckDB, a database tool
58
00:03:20.077 --> 00:03:24.417
that queries data files directly,
reads it with one SQL query.
59
00:03:24.617 --> 00:03:29.177
Collections go by letter here:
Spreadsheets with answers is A,
60
00:03:29.177 --> 00:03:30.877
and Disk images is B.
61
00:03:31.077 --> 00:03:34.057
Again, we keep only sets with answers.
62
00:03:34.257 --> 00:03:36.837
That's 111 material sets.
63
00:03:37.037 --> 00:03:40.257
Now the 6 largest sets
per collection, by units.
64
00:03:40.457 --> 00:03:43.277
Our two picks are here: this one
65
00:03:43.477 --> 00:03:44.537
and this one.
66
00:03:44.737 --> 00:03:46.917
This column is the license tier.
67
00:03:48.437 --> 00:03:52.057
To fetch files, we use the fetch tool
that comes with the index.
68
00:03:52.257 --> 00:03:56.197
uv pulls it from the web and runs it,
with nothing to install.
69
00:03:56.397 --> 00:03:58.257
Both sources go in one command.
70
00:03:58.457 --> 00:04:03.677
We take one unit from each,
with dry run: preview only, no download.
71
00:04:03.877 --> 00:04:07.097
The files not taken are too many to show,
so we filter them out.
72
00:04:07.297 --> 00:04:10.277
You can see the two sources have 55 units.
73
00:04:10.477 --> 00:04:14.477
By default, the tool leaves out
duplicates, benchmark copies, files
74
00:04:14.477 --> 00:04:17.517
not yet compared with the benchmarks,
files we may not ship,
75
00:04:17.517 --> 00:04:21.777
files not in English,
personal data, and files over 20 MB.
76
00:04:21.977 --> 00:04:24.637
This time, all of these are 0.
77
00:04:24.837 --> 00:04:30.057
As asked, it takes 2 units: 8 files,
answers and companion files included.
78
00:04:30.257 --> 00:04:33.797
That looks right,
so we drop dry run and download for real.
79
00:04:34.588 --> 00:04:36.368
All 8 files are OK.
80
00:04:36.568 --> 00:04:39.868
Each one was checked against the size
and checksum in the index.
81
00:04:40.068 --> 00:04:41.668
This exFAT image comes
82
00:04:41.668 --> 00:04:45.688
inside an archive upstream,
and the tool unpacked it for us.
83
00:04:45.888 --> 00:04:47.268
Let's see what we got.
84
00:04:47.721 --> 00:04:52.381
The material goes in folders by domain,
and the answers go apart, under answer.
85
00:04:52.581 --> 00:04:55.221
They're kept apart
because the answers are for grading,
86
00:04:55.221 --> 00:04:57.623
and must never
get into the agent's environment.
87
00:04:58.283 --> 00:05:00.623
The index gives you material and answers.
88
00:05:00.823 --> 00:05:02.923
Building the tasks is up to you.
89
00:05:03.123 --> 00:05:06.343
Here is one way to do it,
in the Harbor task format.
90
00:05:06.543 --> 00:05:09.443
Harbor is a framework
for running tasks like these.
91
00:05:09.643 --> 00:05:11.483
A task is four things:
92
00:05:11.483 --> 00:05:16.303
an instruction, an environment,
a reference solution, and tests.
93
00:05:17.863 --> 00:05:20.043
Let's build the spreadsheet task first.
94
00:05:20.243 --> 00:05:24.263
This Python command and these two paths
are in the example's README.
95
00:05:24.463 --> 00:05:27.923
The paths are the answer file
and the workbook we just fetched.
96
00:05:28.123 --> 00:05:29.823
We open the answer file.
97
00:05:30.381 --> 00:05:39.441
It holds the values of 5 formula cells:
6.5, 6.6, 6.55, 4.9, and 5.8.
98
00:05:39.641 --> 00:05:41.501
Now the workbook itself.
99
00:05:42.194 --> 00:05:46.294
Next to each of the 5 formulas
is its result from the last calculation.
100
00:05:48.294 --> 00:05:51.014
They match the answers exactly,
because the answer file
101
00:05:51.014 --> 00:05:54.954
was taken from these stored results,
saved when the workbook was published.
102
00:05:55.154 --> 00:05:57.414
So the answers are hidden in the workbook.
103
00:05:57.614 --> 00:06:01.394
Give this file to an agent as it is,
and it just reads them off.
104
00:06:01.594 --> 00:06:03.614
The task would be pointless.
105
00:06:04.517 --> 00:06:06.597
So for the copy the agent gets,
106
00:06:06.597 --> 00:06:10.217
we open the workbook
with openpyxl and save it again.
107
00:06:10.417 --> 00:06:11.957
It's just this one line.
108
00:06:12.157 --> 00:06:16.337
openpyxl is a Python library
for reading and writing Excel files.
109
00:06:16.537 --> 00:06:21.037
When it saves, it writes the formulas,
but not the computed results.
110
00:06:24.453 --> 00:06:25.593
Let's look again.
111
00:06:27.437 --> 00:06:29.217
The formulas are all still there.
112
00:06:29.417 --> 00:06:32.877
The stored results have all become None:
they're gone.
113
00:06:33.077 --> 00:06:35.157
The answer file goes
into the tests folder,
114
00:06:35.157 --> 00:06:36.937
and is used only for grading.
115
00:06:37.850 --> 00:06:39.630
The instruction is short.
116
00:06:39.830 --> 00:06:43.290
Work out these 5 cells,
and write them to values.json.
117
00:06:43.490 --> 00:06:46.550
The reference solution
is the standard way to solve the task.
118
00:06:46.750 --> 00:06:48.830
It opens the workbook in LibreOffice,
119
00:06:48.830 --> 00:06:52.050
headless, with no window,
and saves a copy.
120
00:06:52.250 --> 00:06:54.350
Then it reads out the 5 numbers.
121
00:06:54.550 --> 00:06:58.310
The stored results were removed,
so LibreOffice computes these cells
122
00:06:58.310 --> 00:06:59.770
when it opens the file.
123
00:07:00.683 --> 00:07:02.143
The tests are simple.
124
00:07:02.343 --> 00:07:03.383
Each value must match
125
00:07:03.383 --> 00:07:06.603
the answer file within a relative error
of one in a million.
126
00:07:10.256 --> 00:07:14.836
And the tests run in a separate container,
an isolated place to run code.
127
00:07:15.036 --> 00:07:18.176
The answers live
only in the tests container.
128
00:07:18.376 --> 00:07:21.836
Only one file, values.json,
passes between the two.
129
00:07:22.036 --> 00:07:23.976
Both sides are offline, too.
130
00:07:24.176 --> 00:07:25.456
Otherwise, the agent
131
00:07:25.456 --> 00:07:29.396
could download the original file
from upstream and read the stored results.
132
00:07:30.317 --> 00:07:32.669
The second task
has a different kind of answer.
133
00:07:32.857 --> 00:07:37.437
First, we check the image with fsck,
the file system checker.
134
00:07:37.637 --> 00:07:39.657
You can see a string of errors.
135
00:07:39.857 --> 00:07:44.997
A cluster chain is broken: a file's blocks
on the disk no longer link up.
136
00:07:45.197 --> 00:07:46.977
A checksum is wrong, too.
137
00:07:47.177 --> 00:07:48.637
Exit code 4.
138
00:07:48.837 --> 00:07:50.377
It really is broken.
139
00:07:50.577 --> 00:07:52.677
Its answer file holds no value.
140
00:07:52.877 --> 00:07:56.957
It points to the check instead:
the address of the upstream test script,
141
00:07:56.957 --> 00:07:58.257
and its line numbers.
142
00:07:58.457 --> 00:08:01.277
That's how exfatprogs
itself checks a repair.
143
00:08:01.477 --> 00:08:04.537
The image goes into the environment
for the agent to repair.
144
00:08:04.737 --> 00:08:06.191
A copy of the original stays
145
00:08:06.191 --> 00:08:09.891
in tests, so the grading
can confirm the image starts out broken.
146
00:08:10.817 --> 00:08:15.137
So the instruction:
repair it until fsck finds nothing to fix,
147
00:08:15.137 --> 00:08:17.717
save it to this path, and keep its size.
148
00:08:17.917 --> 00:08:19.617
The tests have four checks.
149
00:08:19.845 --> 00:08:23.465
One: the original image
really reports errors.
150
00:08:23.665 --> 00:08:26.885
Two: the repaired image has the same size.
151
00:08:27.085 --> 00:08:31.425
Three: checked again, fsck exits with 0.
152
00:08:31.625 --> 00:08:34.285
Checks one and three follow
the upstream assertions.
153
00:08:34.485 --> 00:08:36.185
Check four is ours.
154
00:08:36.385 --> 00:08:39.105
The three files
and the directory whose names start
155
00:08:39.105 --> 00:08:41.445
with l0 must still be there.
156
00:08:41.645 --> 00:08:45.745
fsck reports no errors on them,
so a real repair keeps them.
157
00:08:45.945 --> 00:08:51.085
Without this check, a freshly formatted
empty image would pass the first three.
158
00:08:52.017 --> 00:08:53.957
The reference solution is short, too.
159
00:08:54.157 --> 00:08:59.057
fsck repairs the image, checks it again,
and copies it to the required path.
160
00:09:00.350 --> 00:09:02.050
Both tasks are built.
161
00:09:02.250 --> 00:09:03.950
Let's look at the folders.
162
00:09:04.150 --> 00:09:06.230
Each task has an instruction,
163
00:09:06.230 --> 00:09:10.010
an environment,
a reference solution, and tests.
164
00:09:10.210 --> 00:09:12.070
That's the Harbor task format.
165
00:09:12.270 --> 00:09:17.150
Following Terminal-Bench 4.0 conventions,
we also gave each task a README,
166
00:09:17.150 --> 00:09:19.890
and a cheat solution,
which we'll use later.
167
00:09:22.509 --> 00:09:24.449
Our Harbor version is pinned.
168
00:09:24.649 --> 00:09:28.329
First, we lint the tasks:
a static check of the folders
169
00:09:28.329 --> 00:09:31.549
and the config against the rules,
without running anything.
170
00:09:31.749 --> 00:09:33.769
We turn on two extra rules.
171
00:09:33.969 --> 00:09:37.569
The agent must not run as root,
and the environment's container image
172
00:09:37.569 --> 00:09:42.069
must be pinned by digest,
that is, by its sha256.
173
00:09:42.269 --> 00:09:44.769
No problems, and the exit code is 0.
174
00:09:45.483 --> 00:09:47.183
Now for the real run.
175
00:09:47.383 --> 00:09:51.643
First, oracle: the agent
that just runs the reference solution.
176
00:09:51.843 --> 00:09:54.723
For each task,
Harbor builds the environment,
177
00:09:54.723 --> 00:09:58.183
runs the solution,
then grades it in a separate container.
178
00:09:58.383 --> 00:10:00.003
We fast-forward this part.
179
00:10:00.203 --> 00:10:02.783
It really took 1 minute 46 seconds.
180
00:10:05.283 --> 00:10:07.943
The mean reward over the two tasks is 1.
181
00:10:08.143 --> 00:10:09.283
Both pass.
182
00:10:11.250 --> 00:10:13.350
But that's not enough.
183
00:10:13.550 --> 00:10:17.150
If the grading is loose,
doing nothing could score too,
184
00:10:17.150 --> 00:10:19.970
and the reference solution passing
would prove nothing.
185
00:10:20.170 --> 00:10:23.550
So next, nop: the agent that does nothing.
186
00:10:26.050 --> 00:10:28.950
Both tasks score 0, as they should.
187
00:10:29.650 --> 00:10:32.470
Scoring 0 for doing nothing
is still not enough.
188
00:10:32.670 --> 00:10:36.050
We also need to see if a shortcut
can fool the grading.
189
00:10:36.250 --> 00:10:38.590
Each task comes with a cheat solution.
190
00:10:38.790 --> 00:10:42.490
For the exFAT task,
it just formats a new image.
191
00:10:42.690 --> 00:10:44.290
For the spreadsheet task,
192
00:10:44.290 --> 00:10:48.230
it downloads the original workbook,
and copies its stored results.
193
00:10:48.430 --> 00:10:51.710
Following the example's README,
we put the cheat solution
194
00:10:51.710 --> 00:10:54.850
where the reference solution goes,
and run oracle again.
195
00:10:58.891 --> 00:11:01.151
Both tasks score 0 again.
196
00:11:02.217 --> 00:11:03.357
The rewards:
197
00:11:03.557 --> 00:11:05.957
both tasks get 1
with the reference solution,
198
00:11:05.957 --> 00:11:08.217
and 0 in the other two runs.
199
00:11:08.417 --> 00:11:09.957
For the spreadsheet task,
200
00:11:10.497 --> 00:11:13.397
the 5 computed numbers
match the answers exactly.
201
00:11:13.597 --> 00:11:17.457
The nop and cheat runs left
no output file, so they fail right away.
202
00:11:17.657 --> 00:11:21.357
The cheat needs the network to download,
but the task is offline.
203
00:11:21.557 --> 00:11:23.097
For the exFAT task,
204
00:11:23.742 --> 00:11:26.162
the reference solution
passes all four checks.
205
00:11:26.362 --> 00:11:28.622
The nop run fails the last three.
206
00:11:28.822 --> 00:11:32.522
The cheat's new image passes
the first three, but fails the fourth.
207
00:11:32.722 --> 00:11:34.582
That's why we added check four.
208
00:11:40.450 --> 00:11:43.590
So we went from a public data page
to two tasks.
209
00:11:43.790 --> 00:11:47.810
The reference solution passes,
and doing nothing or cheating fails.
210
00:11:48.010 --> 00:11:50.910
That's the minimum a task
needs before you train on it.
211
00:11:51.110 --> 00:11:56.790
To recap: pick material, fetch files,
take the answers out of the environment,
212
00:11:56.790 --> 00:12:02.390
write the instruction and the tests,
lint them, and finally run oracle, nop,
213
00:12:02.390 --> 00:12:04.330
and the cheat once each.
214
00:12:04.530 --> 00:12:07.670
With another material set,
the steps are the same.
215
00:12:07.870 --> 00:12:11.090
The data page and the example
are linked on screen.
216
00:12:11.290 --> 00:12:15.550
Both tasks and every command
are in the example, so give it a try.