Datasets:
Update README.md
Browse files
README.md
CHANGED
@@ -314,45 +314,44 @@ configs:
|
|
314 |
task_categories:
|
315 |
- text-classification
|
316 |
- text-generation
|
|
|
|
|
|
|
|
|
|
|
|
|
317 |
language:
|
318 |
- it
|
319 |
size_categories:
|
320 |
- 100B<n<1T
|
321 |
---
|
322 |
-
|
323 |
-
NEXT FEW DAYS AND FILLED WITH MANY OTHER
|
324 |
-
INFORMATIONS AND DETAILED STATISTICS
|
325 |
-
|
326 |
![image/jpeg](https://cdn-uploads.huggingface.co/production/uploads/65c27751d2fbc4e846637421/c726oXPuXRbFrLiAZqKs5.jpeg)
|
327 |
|
328 |
-
|
329 |
-
|
330 |
-
|
331 |
-
|
332 |
-
|
333 |
-
|
334 |
-
|
335 |
-
|
336 |
-
|
337 |
-
the
|
338 |
-
|
339 |
-
|
340 |
-
libraries such as BeautifulSoup and Selenium; the scripts were mostly
|
341 |
-
|
342 |
-
|
343 |
-
|
344 |
-
|
345 |
-
|
346 |
-
|
347 |
-
|
348 |
-
|
349 |
-
|
350 |
-
|
351 |
-
|
352 |
-
other possible purposes. The Tiktoken BPE tokenizer with the
|
353 |
-
cl100k\_base model \[2\] was used for tokenization. This dataset is
|
354 |
-
composed of several sub-datasets, each with different types of data and
|
355 |
-
goals.
|
356 |
|
357 |
**Conversational (\~ 85 Billions tokens):**
|
358 |
|
@@ -367,10 +366,7 @@ This split contains 19.395.579.455 tokens. Texts were not checked for
|
|
367 |
language, but it is a safe assumption that most of the text contained is
|
368 |
in Italian as the selected Usenet hierarchies target only Italian users.
|
369 |
|
370 |
-
|
371 |
-
here are general stats about this part of the dataset:
|
372 |
-
|
373 |
-
|
374 |
|
375 |
\"chars": 59389804791,
|
376 |
|
@@ -378,27 +374,40 @@ here are general stats about this part of the dataset:
|
|
378 |
|
379 |
\"sentences": 519535427,
|
380 |
|
381 |
-
\"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
382 |
|
383 |
-
\"thread": 14521548,
|
384 |
|
385 |
|
386 |
83GB of JSONL file before the conversion to HuggingFace dataset
|
387 |
|
|
|
|
|
388 |
**Forum**
|
389 |
|
390 |
The second part of the project is the one that produced the largest
|
391 |
-
amount of data
|
392 |
different platforms (phpBB, vBulletin, Simple Machines, Invision, Snitz,
|
393 |
XenForo\...) was created using both manual and semi-automatic web
|
394 |
searches. Then, for each forum, a generic script (forum\_scraper.py)
|
395 |
using Python3 and BeautifulSoup was adapted to fit the characteristics
|
396 |
of the forum (such as correct div classes for the different fields and
|
397 |
multiple page mechanisms). Then, the script ran over the entire range of
|
398 |
-
available pages and output a JSONL file with one post per line.
|
399 |
-
statistics, already computed, will follow very soon. For now, here are
|
400 |
-
general stats about this part of the dataset:
|
401 |
|
|
|
402 |
{
|
403 |
|
404 |
\"chars": 199436329709,
|
@@ -415,6 +424,18 @@ general stats about this part of the dataset:
|
|
415 |
|
416 |
}
|
417 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
418 |
303GB of JSONL files before the conversion to HuggingFace dataset.
|
419 |
|
420 |
Regarding multimodality, in short: this feature is not very well
|
@@ -465,28 +486,20 @@ as they would have been publicly available on the web.
|
|
465 |
This split of the dataset contains articles published as Open Access
|
466 |
using the platform OJS. It comprised mainly academic journals from
|
467 |
Italian universities, so it can be considered as a very high-quality
|
468 |
-
dataset. All the articles are published with Creative Commons licenses,
|
469 |
and the license used for the single article can be retrieved from the
|
470 |
metadata.
|
|
|
|
|
|
|
|
|
|
|
|
|
471 |
|
472 |
**Blogs**
|
473 |
|
474 |
-
This resource was gathered by scraping data from blogs written in
|
475 |
-
Italian.
|
476 |
-
left-wing activism, in order to help another person for his research
|
477 |
-
project, that it is still work in progress. The list of these blog was
|
478 |
-
obtained on a blog aggregator. The blogs that fall under this category
|
479 |
-
are labelled with the category "pol/ant" (Poltics/Antagonism). Because
|
480 |
-
from a quick analysis it seems that data coming from the "forum"
|
481 |
-
category are mainly biased toward right political stances (data about
|
482 |
-
this statement will follow in the next weeks), it could be useful to
|
483 |
-
integrate these data in a general language-modelling task in the optic
|
484 |
-
of the "Perspectivist Data Manifesto" \[1\]. The other two categories
|
485 |
-
are "let/litblog", containing blogs about literature (the list was
|
486 |
-
obtained from another aggregator) and "inf/linux", a very small category
|
487 |
-
containing blog posts from Italian Linux User Groups. The rest of the
|
488 |
-
data is not categorized. Here a breakdown of number of tokens per
|
489 |
-
category:
|
490 |
|
491 |
This sub-project started with the goal of collecting only blogs released
|
492 |
under Public Domain or Creative Commons license. However, due do the
|
@@ -501,6 +514,21 @@ is fine under Fair-Use for research purposes but the possibility of
|
|
501 |
usage should be checked by whom wants to use this dataset for other
|
502 |
purposes, especially for commercial purposes.
|
503 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
504 |
This resource can be considered as a "medium-high" quality dataset,
|
505 |
because it mostly contain blogs post, often from good sources with very
|
506 |
informative content. It is not possible to guarantee a total absence of
|
@@ -508,16 +536,28 @@ undesired content inside the resource, but this, depending from the use
|
|
508 |
case, probably constitutes a minority.
|
509 |
|
510 |
As for the Conversational data split, also this split is diachronically
|
511 |
-
annotated so it could be used for
|
512 |
|
513 |
Finally, the blog split contains also an annotation for the language
|
514 |
used, as identified by the FastText library.
|
515 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
516 |
**Wikimedia**
|
517 |
|
518 |
This split doesn't need many explanation as it is simply a dump of
|
519 |
wikimedia resources in Italian (Wikipedia, Wikibooks, Wikinews,
|
520 |
-
Wikiquote, Wikisource, Wikiversity, Wikivoyage and Wiktionary). It can
|
521 |
be very important to include this resource in the training data of a
|
522 |
language model because it contains information, presented in a mostly
|
523 |
neutral language, about many possible subjects and topics that are not
|
@@ -532,6 +572,11 @@ included in this split are: eml (emilian e rumagno) ,fur (furlan) ,la
|
|
532 |
(sicilianu) ,sc (sardu) and vec (veneto). Using this data, depending
|
533 |
from the goal of the project, could produce very interesting results.
|
534 |
|
|
|
|
|
|
|
|
|
|
|
535 |
**Books**
|
536 |
|
537 |
This collection contains mainly the books coming from LiberLiber's
|
@@ -544,7 +589,14 @@ of Italian culture.
|
|
544 |
The collection contains also a smaller amount of data coming from other
|
545 |
sources, such as the Creative Commons licensed school books of
|
546 |
"Matematicamente" \[3\] and Oilproject-Weschool \[4\] as well as some
|
547 |
-
other CC and PD
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
548 |
|
549 |
**Websites**
|
550 |
|
@@ -556,7 +608,7 @@ contains many official documents about Mafia persecution in Italy,
|
|
556 |
Peacelink \[6\], an historical Italian website about peace activism and
|
557 |
HomoLaicus \[7\] a big collection of texts about various topics (mainly
|
558 |
history and politics) released under a CC license. Also other smaller
|
559 |
-
and randomly selected websites are included in this collection. This
|
560 |
section has to be considered experimental for two reasons: (1) It
|
561 |
containly only a very small subset of the entire high-quality Italian
|
562 |
web landscape and it could be increased and improved "ad libitum" (2) It
|
@@ -566,6 +618,10 @@ that we will discuss in the appropriate section.
|
|
566 |
Despite these two point, users are encouraged to use this section as it
|
567 |
is composed of medium-high and high quality contents.
|
568 |
|
|
|
|
|
|
|
|
|
569 |
**Reddit**
|
570 |
|
571 |
It contains a small subsets (4192672 messages) of conversations in some
|
@@ -574,7 +630,37 @@ Italian subreddits.
|
|
574 |
**Italatex**
|
575 |
Still work in progress. A collection of materials written in LaTeX.
|
576 |
|
577 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
578 |
|
579 |
The presence of duplicate text can be, depending from the use cases, a
|
580 |
big problem for several machine learning tasks. I tried to avoid as much
|
@@ -619,16 +705,65 @@ in the form of 1) header of the website 2) list of links 3) footer of
|
|
619 |
the website. All the HTML was converted using HTML2TEXT so it should not
|
620 |
contain html code.
|
621 |
|
|
|
622 |
|
623 |
-
|
624 |
-
*Work in progress; this will contain statistics of tokens, chars and sentences lenght for each diachronic resource (usenet newsgroup, post, blog) for each month for each year*
|
625 |
|
626 |
-
|
627 |
-
|
628 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
629 |
|
630 |
\* \[1\] <https://pdai.info/>
|
631 |
|
632 |
\* \[2\] https://github.com/openai/tiktoken
|
633 |
|
634 |
-
\* \[3\] <https://xmau.com/usenet/>
|
|
|
|
|
|
|
|
314 |
task_categories:
|
315 |
- text-classification
|
316 |
- text-generation
|
317 |
+
annotations_creators:
|
318 |
+
- no-annotation
|
319 |
+
multilinguality:
|
320 |
+
- monolingual
|
321 |
+
task_ids:
|
322 |
+
- language-modeling
|
323 |
language:
|
324 |
- it
|
325 |
size_categories:
|
326 |
- 100B<n<1T
|
327 |
---
|
328 |
+
# Dataset Card for Testimole -- A multi-billion tokens Italian text corpus
|
|
|
|
|
|
|
329 |
![image/jpeg](https://cdn-uploads.huggingface.co/production/uploads/65c27751d2fbc4e846637421/c726oXPuXRbFrLiAZqKs5.jpeg)
|
330 |
|
331 |
+
Testimole is a large linguistic resource for Italian obtained through a massive web scraping effort. As of June 2024, it is one of the largest datasets for the Italian language, if not the largest, publicly available, consisting of almost 100B tokens counted with the Tiktoken cl100k BPE tokenizer. It consists mainly of conversational data (Italian Usenet hierarchies, Italian message boards, Italian subreddits) but also contains other textual data derived from blogs, wikis, websites, and academic journals. Each data source is separated into a different dataset split.
|
332 |
+
|
333 |
+
Testimole is a wordplay combining "Testi" (texts) and "Mole." "Mole" refers to one of the most famous monuments of Torino, the [Mole Antonelliana](https://en.wikipedia.org/wiki/Mole_Antonelliana), where this dataset was conceived and built. Moreover, "mole" means "mass" or "bulk" in Italian, highlighting the large size of this dataset. Testimole is also similar to the word "Testimone" (witness), emphasizing the diachronic quality of the data, thus being a witness of the passage of time in the Italian webosphere.
|
334 |
+
|
335 |
+
|
336 |
+
## Dataset Details
|
337 |
+
|
338 |
+
### Dataset Description
|
339 |
+
|
340 |
+
The goal of this work is to create a huge linguistic resource for the Italian language that can be used for several NLP applications, including but not limited to language modelling. The dataset is the result of a massive web scraping effort going on from February 2024 to May 2024, so the resources have a cut-off date comprised within this time span.
|
341 |
+
There is a project to further expand the dataset, as explained in the "Future Plans" section.
|
342 |
+
|
343 |
+
To create the dataset, I developed several scripts using Python3 and libraries such as BeautifulSoup and Selenium; the scripts were mostly written and executed manually, making it an extremely time-consuming project. The texts span different topics and periods, containing several divergent opinions and beliefs, in accordance with the main ideas of the "Perspective Data Manifesto" [1]. It is important to note that these data alone are not enough to train an Italian large language model from scratch, mainly not due to the size of the data but because, even if they span different topics, they are far from covering the broad range of subjects, information, culture, and techniques required to train a state-of-the-art model. Also, as will be better pointed out later, while it is safe to use these data under Fair Use for research purposes, users must investigate potential copyright infringement for other possible purposes. The Tiktoken BPE tokenizer with the cl100k_base model [2] was used for tokenization. This dataset is composed of several sub-datasets, each with different types of data and goals.
|
344 |
+
|
345 |
+
## Uses
|
346 |
+
|
347 |
+
Because this dataset consists of a large amount of texts in the Italian language, it can be used for all Natural Language applications that seek to improve support for Italian in a multilingual context and require data for training. This includes, but is not limited to, training large language models. Other possible uses are sentiment analysis, diachronic data classification (as the majority of the data is date-tagged), and text classification. Researchers are invited to annotate even small parts of this dataset. In such cases, the data could be used for other tasks as well, such as Named Entity Recognition (NER), Part-of-Speech (POS) tagging, information retrieval, summarization, and more. This versatility makes the dataset a valuable resource for various NLP projects and research endeavors.
|
348 |
+
|
349 |
+
### Out-of-Scope Use
|
350 |
+
|
351 |
+
By downloading this dataset, users agree not to attempt to identify specific users. This includes refraining from cross-referencing the dataset with other information to ascertain personal identities.
|
352 |
+
|
353 |
+
## Dataset Structure
|
354 |
+
|
|
|
|
|
|
|
|
|
355 |
|
356 |
**Conversational (\~ 85 Billions tokens):**
|
357 |
|
|
|
366 |
language, but it is a safe assumption that most of the text contained is
|
367 |
in Italian as the selected Usenet hierarchies target only Italian users.
|
368 |
|
369 |
+
*General stats:*
|
|
|
|
|
|
|
370 |
|
371 |
\"chars": 59389804791,
|
372 |
|
|
|
374 |
|
375 |
\"sentences": 519535427,
|
376 |
|
377 |
+
\"posts": 89499446,
|
378 |
+
|
379 |
+
\"threads": 14521548
|
380 |
+
|
381 |
+
|
382 |
+
*Columns of the dataset*
|
383 |
+
* Title: The original title of the thread
|
384 |
+
* Author: Author of the post
|
385 |
+
* ID: a unique identifier of the post for the specific newsgroup
|
386 |
+
* Progressive_id: the progressive id of the single message in the thread
|
387 |
+
* Timestamp: the time and data of creation of the post, in ISO-8601 format
|
388 |
+
* Newsgroup: the name of the newsgroup in which the post belong
|
389 |
+
* Original_url: the URL of the thread
|
390 |
+
* Text: the text of the message
|
391 |
|
|
|
392 |
|
393 |
|
394 |
83GB of JSONL file before the conversion to HuggingFace dataset
|
395 |
|
396 |
+
|
397 |
+
|
398 |
**Forum**
|
399 |
|
400 |
The second part of the project is the one that produced the largest
|
401 |
+
amount of data (62.415.825.978 tokens) A list of Italian message boards based on
|
402 |
different platforms (phpBB, vBulletin, Simple Machines, Invision, Snitz,
|
403 |
XenForo\...) was created using both manual and semi-automatic web
|
404 |
searches. Then, for each forum, a generic script (forum\_scraper.py)
|
405 |
using Python3 and BeautifulSoup was adapted to fit the characteristics
|
406 |
of the forum (such as correct div classes for the different fields and
|
407 |
multiple page mechanisms). Then, the script ran over the entire range of
|
408 |
+
available pages and output a JSONL file with one post per line.
|
|
|
|
|
409 |
|
410 |
+
*General stats:*
|
411 |
{
|
412 |
|
413 |
\"chars": 199436329709,
|
|
|
424 |
|
425 |
}
|
426 |
|
427 |
+
*Columns of the dataset*
|
428 |
+
* Title: The original title of the thread
|
429 |
+
* Author: Author of the post
|
430 |
+
* post_id: a unique identifier of the post for the specific forum
|
431 |
+
* Progressive_id: the progressive id of the single message in the thread
|
432 |
+
* Timestamp: the time and data of creation of the post. In the majority of cases, it is in ISO-8601 format but sometime it could be not converted to ISO-8601 and so being in other formats (a good future work is to convert everything to ISO-8601). In rare cases, it is set to None.
|
433 |
+
* Forum: the name of the forum. If the forum belongs to the Forumfree or Forumcommunity circuit, the name of the circuit is appended to the name of the forum. There are cases of forums belonging to the Forumfree circuit where Forumfree is not appended. This should be fixed in a future release.
|
434 |
+
* Text: the text of the message
|
435 |
+
* image_list: experimental multimodal support
|
436 |
+
* image_file: experimental multimodal support
|
437 |
+
|
438 |
+
|
439 |
303GB of JSONL files before the conversion to HuggingFace dataset.
|
440 |
|
441 |
Regarding multimodality, in short: this feature is not very well
|
|
|
486 |
This split of the dataset contains articles published as Open Access
|
487 |
using the platform OJS. It comprised mainly academic journals from
|
488 |
Italian universities, so it can be considered as a very high-quality
|
489 |
+
dataset and not problematic regarding biases, apart from very generic biases that may be present in the Italian language in itself or in Academia environments. All the articles are published with Creative Commons licenses,
|
490 |
and the license used for the single article can be retrieved from the
|
491 |
metadata.
|
492 |
+
*Columns of the dataset*
|
493 |
+
* Journal:
|
494 |
+
* url:
|
495 |
+
* metadata:
|
496 |
+
* text:
|
497 |
+
* platform:
|
498 |
|
499 |
**Blogs**
|
500 |
|
501 |
+
This resource was gathered by scraping data from blogs and on-line newspapers written in
|
502 |
+
Italian.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
503 |
|
504 |
This sub-project started with the goal of collecting only blogs released
|
505 |
under Public Domain or Creative Commons license. However, due do the
|
|
|
514 |
usage should be checked by whom wants to use this dataset for other
|
515 |
purposes, especially for commercial purposes.
|
516 |
|
517 |
+
The project started with a collection of blogs regarding
|
518 |
+
left-wing activism, in order to help another person for his research
|
519 |
+
project, that it is still work in progress. The list of these blog was
|
520 |
+
obtained on a blog aggregator. The blogs that fall under this category
|
521 |
+
are labelled with the category "pol/ant" (Poltics/Antagonism). Because
|
522 |
+
from a quick analysis it seems that data coming from the "forum"
|
523 |
+
category are mainly biased toward right political stances (data about
|
524 |
+
this statement will follow in the next weeks), it could be useful to
|
525 |
+
integrate these data in a general language-modelling task in the optic
|
526 |
+
of the "Perspectivist Data Manifesto" \[1\]. The other two categories
|
527 |
+
are "let/litblog", containing blogs about literature (the list was
|
528 |
+
obtained from another aggregator) and "inf/linux", a very small category
|
529 |
+
containing blog posts from Italian Linux User Groups. The rest of the
|
530 |
+
data, which account for the majority of tokens, is not categorized.
|
531 |
+
|
532 |
This resource can be considered as a "medium-high" quality dataset,
|
533 |
because it mostly contain blogs post, often from good sources with very
|
534 |
informative content. It is not possible to guarantee a total absence of
|
|
|
536 |
case, probably constitutes a minority.
|
537 |
|
538 |
As for the Conversational data split, also this split is diachronically
|
539 |
+
annotated so it could be used for diachronic analysis of language and topics too.
|
540 |
|
541 |
Finally, the blog split contains also an annotation for the language
|
542 |
used, as identified by the FastText library.
|
543 |
|
544 |
+
*Columns of the dataset*
|
545 |
+
* title: The title of the article/post
|
546 |
+
* name: The name of the blog
|
547 |
+
* author: The author of the article/post, if available
|
548 |
+
* date: The date of the article/post in ISO-8601, if available if not None
|
549 |
+
* url: The original URL
|
550 |
+
* text: The text of the article/post
|
551 |
+
* category: The category of the blog. Only a few blogs are annotated for category up to now.
|
552 |
+
* license_guess: A guess of the original license of the blog made by an automated and non-perfect script
|
553 |
+
* fasttext_langid: The most probable language as identified by fasttext
|
554 |
+
* fasttext_langprob: The probability of the most probable language as identified by fasttext
|
555 |
+
|
556 |
**Wikimedia**
|
557 |
|
558 |
This split doesn't need many explanation as it is simply a dump of
|
559 |
wikimedia resources in Italian (Wikipedia, Wikibooks, Wikinews,
|
560 |
+
Wikiquote, Wikisource, Wikiversity, Wikivoyage and Wiktionary) as of May 2024. It can
|
561 |
be very important to include this resource in the training data of a
|
562 |
language model because it contains information, presented in a mostly
|
563 |
neutral language, about many possible subjects and topics that are not
|
|
|
572 |
(sicilianu) ,sc (sardu) and vec (veneto). Using this data, depending
|
573 |
from the goal of the project, could produce very interesting results.
|
574 |
|
575 |
+
*Columns of the dataset*
|
576 |
+
* title
|
577 |
+
* text
|
578 |
+
* wiki
|
579 |
+
|
580 |
**Books**
|
581 |
|
582 |
This collection contains mainly the books coming from LiberLiber's
|
|
|
589 |
The collection contains also a smaller amount of data coming from other
|
590 |
sources, such as the Creative Commons licensed school books of
|
591 |
"Matematicamente" \[3\] and Oilproject-Weschool \[4\] as well as some
|
592 |
+
other CC and PD licenses book found online.
|
593 |
+
|
594 |
+
*Columns of the dataset*
|
595 |
+
* title
|
596 |
+
* author
|
597 |
+
* url
|
598 |
+
* text
|
599 |
+
|
600 |
|
601 |
**Websites**
|
602 |
|
|
|
608 |
Peacelink \[6\], an historical Italian website about peace activism and
|
609 |
HomoLaicus \[7\] a big collection of texts about various topics (mainly
|
610 |
history and politics) released under a CC license. Also other smaller
|
611 |
+
and randomly selected (but filtered for quality) websites are included in this collection. This
|
612 |
section has to be considered experimental for two reasons: (1) It
|
613 |
containly only a very small subset of the entire high-quality Italian
|
614 |
web landscape and it could be increased and improved "ad libitum" (2) It
|
|
|
618 |
Despite these two point, users are encouraged to use this section as it
|
619 |
is composed of medium-high and high quality contents.
|
620 |
|
621 |
+
*Columns of the dataset*
|
622 |
+
* url
|
623 |
+
* text
|
624 |
+
|
625 |
**Reddit**
|
626 |
|
627 |
It contains a small subsets (4192672 messages) of conversations in some
|
|
|
630 |
**Italatex**
|
631 |
Still work in progress. A collection of materials written in LaTeX.
|
632 |
|
633 |
+
## Dataset Creation
|
634 |
+
|
635 |
+
### Curation Rationale
|
636 |
+
Multilinguality is one of the main challenges for the new AI and NLP revolution that is taking place in the 2020s. Until now, the most advanced models are mostly trained on English or a few other languages, creating a dangerous gap for people speaking other languages (that is, 81.2% of the world population, according to the CIA Factbook of 2022) in accessing these new advanced instruments. Translations from English are not enough to capture the cultural differences of peoples that do not belong to the Anglo-American culture. Thus, it is important for a general model to be truly inclusive by including data that can capture different views of the world and uses of language. These data are not only useful for modeling the Italian language itself but also for gaining insights into Italian culture and the way in which Italian-speaking people engage with various topics.
|
637 |
+
### Source Data
|
638 |
+
* Usenet
|
639 |
+
* Message boards
|
640 |
+
* Blogs
|
641 |
+
* Websites
|
642 |
+
* Open Journal System platforms hosted by Italian universities or included in DOAJ
|
643 |
+
* LiberLiber
|
644 |
+
* Wikimedia
|
645 |
+
|
646 |
+
#### Data Collection and Processing
|
647 |
+
|
648 |
+
The dataset is the result of a massive web scraping effort that was carried out using manually created Python3 scripts using libraries such as BeautifulSoup for HTML parsing and Selenium in the few cases in which Javascript support or browser automatization was required. I have created blueprints of the script, such a generic "forum scraper" or "blog scraper" script but then I had to adapt them almost manually for each resource included in the dataset. Some resources were sharing the same technical platform, so it was trivial to adapt the script, in other case a significant reverse-engineering effort was required.
|
649 |
+
The scraping took place on very simple resources: a very old unused 2006 Sony Vaio laptop with an Intel Core2Duo processor connected to a domestic FTTC connection was enough for the majority of websites, while in other cases other resources were rented or borrowed in order to have a speed-up or to aggregate and analyse the entire collection of data. Using such a simple setup was also a way to have a "natural" anti-overload system. Because many web scraping instances were running in parallel, websites were not loaded so much and often timeouts were implemented in order to protect smaller servers. It never happened that a website was slowed down due to this scraping process, that was crafted to be as gentle and slow as possible.
|
650 |
+
Considering an average power consumption of the laptop of 40W and 100 days of running, circa 96Kwh were used to power the laptop. The laptop was plugged in the Italian-Centre electricity zone, with an average electricity/co2 ratio of 250g per KW with more than 60% of power coming from renewable sources. This means that the laptop Co2 emissions were circa 24KG of Co2, equivalent to a short 150km trip on a small car with emissions of 160g/km.
|
651 |
+
All the data were collected in a JSONL format and then merged, cleaned, analyzed and converted to an Hugging Face dataset using an HPC resource that was gently provided to the author.
|
652 |
+
The vast majority of data coming from forums undergone a process of deduplication in order to avoid the case of having two instances of the same message.
|
653 |
+
|
654 |
+
|
655 |
+
#### Who are the source data producers?
|
656 |
+
|
657 |
+
Data is produced by users of the Italian Internet mostly between 1995 and 2024. This resource also contains texts produced before 1995, such as the content of public domain books written by authors from any historical period.
|
658 |
+
|
659 |
+
## Recommendations
|
660 |
+
|
661 |
+
This dataset could be used along with other Italian natural language resources. A very good list of them is available at the address [https://huggingface.co/collections/gsarti/italian-nlp-resources-64fc606927fb3a92e9ea72f2]. For example, gsarti/clean_mc4_it [https://huggingface.co/datasets/gsarti/clean_mc4_it], being the biggest as-to-date cleaned version of Common Crawl for Italian, could be used to increase the variety of the data for the training of a Large Language Model.
|
662 |
+
|
663 |
+
### Deduplication
|
664 |
|
665 |
The presence of duplicate text can be, depending from the use cases, a
|
666 |
big problem for several machine learning tasks. I tried to avoid as much
|
|
|
705 |
the website. All the HTML was converted using HTML2TEXT so it should not
|
706 |
contain html code.
|
707 |
|
708 |
+
## Citation
|
709 |
|
710 |
+
### Citation Information
|
|
|
711 |
|
712 |
+
```
|
713 |
+
@software{testimole,
|
714 |
+
author = {Rinaldi},
|
715 |
+
title = {TestiMole},
|
716 |
+
month = May,
|
717 |
+
year = 2024,
|
718 |
+
url = {https://huggingface.co/datasets/mrinaldi/TestiMole}
|
719 |
+
}
|
720 |
+
|
721 |
+
```
|
722 |
+
## Future work
|
723 |
+
The dataset could be enhanced in several ways:
|
724 |
+
* Increasing the amount of data (scraping): this could be done both by recycling the same scripts for other forums, blogs and websites or by writing new scraping scripts. It is important to understand that, even if the dataset is big, it only capture a small amount of the entire Italian webosphere;
|
725 |
+
* Increasing the amount of data (not by scraping): we have projects in mind to increase the dataset with high quality contents coming from different sources, expecially from Italian Universities;
|
726 |
+
* Cleaning the data: deduplication, as explained in the appropriate section, is probably the top-priority work that should be done on this dataset;
|
727 |
+
* Much more :) You are warmly invited to collaborate with me in this effort.
|
728 |
+
|
729 |
+
## Statistics
|
730 |
+
More statistics will be added in the near future. In the "asset" directory, JSONL files contain precomputed token counts for each subcategory (e.g., single forum, newsgroup, or blog), allowing anyone interested to easily craft more detailed statistics.
|
731 |
+
|
732 |
+
# Conversational aggregated tokens per year:
|
733 |
+
|
734 |
+
![image/png](https://cdn-uploads.huggingface.co/production/uploads/65c27751d2fbc4e846637421/VTyRGgjgyOFZkCp6xkowA.png)
|
735 |
+
|
736 |
+
# Forum tokens per year:
|
737 |
+
|
738 |
+
![image/png](https://cdn-uploads.huggingface.co/production/uploads/65c27751d2fbc4e846637421/ThgfDRWxfu2JqkPbu6VOi.png)
|
739 |
+
|
740 |
+
# Usenet tokens per year:
|
741 |
+
|
742 |
+
![image/png](https://cdn-uploads.huggingface.co/production/uploads/65c27751d2fbc4e846637421/PzBZb5lhDMX05QQ8US7mn.png)
|
743 |
+
|
744 |
+
# Usenet hierarchies breakdown:
|
745 |
+
|
746 |
+
![image/png](https://cdn-uploads.huggingface.co/production/uploads/65c27751d2fbc4e846637421/_uSTThPiuY_upujSyIZNG.png)
|
747 |
+
|
748 |
+
# Usenet and Forum in diachronic perspective:
|
749 |
+
|
750 |
+
![image/png](https://cdn-uploads.huggingface.co/production/uploads/65c27751d2fbc4e846637421/D5PGNdX9t1aWsBhhaJgnO.png)
|
751 |
+
|
752 |
+
# Blogs tokens per year:
|
753 |
+
|
754 |
+
![image/png](https://cdn-uploads.huggingface.co/production/uploads/65c27751d2fbc4e846637421/T_7-Le3o-hQFK0QSfegMG.png)
|
755 |
+
|
756 |
+
## Contributions
|
757 |
+
Special thanks to [Viviana Patti](https://www.unito.it/persone/vpatti) and [Valerio Basile](https://valeriobasile.github.io/), professors at the Computer Science Department, NLP group, at the University of Turin, Italy who are significantly supporting me and my projects. Grazie mille :)
|
758 |
+
Thanks also to [ruggsea](https://huggingface.co/ruggsea), who helped in the first stage of the creation of the Usenet dataset by converting the first JSONL files to parquet and giving some resources to download part of the dataset. Thanks to the entire [mii-community](https://huggingface.co/mii-community) who supported and expressed interest in the project. Thanks to Luisa for plugging the old laptop, giving me SSH access and reboot it in cases such as power surges or crashes.
|
759 |
+
|
760 |
+
## References (partial)
|
761 |
|
762 |
\* \[1\] <https://pdai.info/>
|
763 |
|
764 |
\* \[2\] https://github.com/openai/tiktoken
|
765 |
|
766 |
+
\* \[3\] <https://xmau.com/usenet/>
|
767 |
+
|
768 |
+
|
769 |
+
|