Datasets:
Domain and provenance annotation
Hello, I believe the usability of the minipile would be greatly increased if the domain or original corpus of each example were available. This would open up the possibility of conducting analyses on generalization and domain effects.
For instance, one could analyze the effects of including Github code examples when attempting to understand dialogues and books. In particular, determining whether this transfer is mostly syntactic in nature, or also semantic, etc.
Agreed, this is a great suggestion, thank you very much! I believe this could be achieved by simple string matching with the original corpus, including metadata. I used this dump here, which didn't include the metadata; otherwise, I would have already added it.
I will try to look into this in the next few days, but I cannot promise an update soon due to other current commitments on my side.
Hi
@JeanKaddour
,
I second
@haukur
's comment. For a project I am working on, it would be great to be able to control the distribution of domains.
Thanks for publishing this great dataset!
Hi, I would be also interested in the domains/original corpus of each sample. Because of that I would be curious if there are updates. :-)
I retrieved the original subcorpus identifier by looking up each example in the original pile, I'll see if I can send a pull request to include it here.
In the meanwhile, these are the global statistics:
| Subcorpus | train [tokens] | train [docs] | valid [docs] | test [docs] |
|-------------------+----------------+--------------+--------------+-------------|
| ArXiv | 72220678 | 11430 | 5 | 138 |
| BookCorpus2 | 3892604 | 69 | | 1 |
| Books3 | 51874282 | 597 | | 6 |
| DM Mathematics | 16343145 | 9314 | 10 | 109 |
| Enron Emails | 514761 | 1695 | | 19 |
| EuroParl | 341702 | 32 | | |
| FreeLaw | 65098428 | 22945 | | 246 |
| Github | 58234990 | 118608 | 52 | |
| Gutenberg (PG-19) | 3770325 | 59 | | 2 |
| HackerNews | 4035984 | 4709 | 2 | 38 |
| NIH ExPorter | 2306630 | 7440 | 3 | 57 |
| OpenSubtitles | 12359258 | 2400 | 1 | 35 |
| OpenWebText2 | 77788242 | 113924 | 59 | 1123 |
| PhilPapers | 2518405 | 276 | | 1 |
| Pile-CC | 265299750 | 328387 | 182 | 3263 |
| PubMed Abstracts | 25421463 | 132144 | 59 | 1317 |
| PubMedCentral | 108849098 | 24405 | 13 | 277 |
| StackExchange | 40478243 | 137446 | 72 | 1376 |
| Ubuntu IRC | 29391239 | 93 | | |
| USPTO Backgrounds | 4105644 | 44897 | 19 | 431 |
| Wikipedia (en) | 19871654 | 38169 | 14 | 412 |
| YoutubeSubtitles | 2449108 | 961 | | 12 |
|-------------------+----------------+--------------+--------------+-------------|
| Size in MB | - | 5847 | 2.8 | 58 |
Hi
@haukur
I just wanted to ask you if you need help, etc., to get the PR set up with the additional meta data info.
I'm asking because I'm about to tokenize the MiniPile but with the subcorpus info I would exclude some subsets that are not needed for my setup.
