[00:00.000 --> 00:07.440]  So let's dive right into it.
[00:07.440 --> 00:11.840]  LibreTranslate is a software that's a bit like Google Translate,
[00:11.840 --> 00:15.200]  but open-source. It is AGPL3 license,
[00:15.200 --> 00:17.040]  so it's a strongly open-source.
[00:17.040 --> 00:19.960]  In fact, we're going to keep it that way forever,
[00:19.960 --> 00:22.840]  and let's you do natural language translation.
[00:22.840 --> 00:25.440]  It runs on your computer.
[00:25.440 --> 00:27.240]  This is one of the goals of the project.
[00:27.240 --> 00:29.440]  There are several other projects in
[00:29.440 --> 00:31.520]  the open-source realm that have aimed to
[00:31.520 --> 00:34.800]  provide natural language translation,
[00:34.800 --> 00:37.040]  except sometimes that they require
[00:37.040 --> 00:39.160]  a very large servers or a lot of memory,
[00:39.160 --> 00:41.440]  and our goal is to have this running on
[00:41.440 --> 00:42.840]  something as low as a Raspberry Pi.
[00:42.840 --> 00:45.560]  So that is very important to the project.
[00:45.560 --> 00:49.440]  The program has lots of clients and integrations.
[00:49.440 --> 00:52.920]  We'll cover some of those in the upcoming slides.
[00:52.920 --> 00:55.840]  Like many projects, it's available on GitHub,
[00:55.840 --> 00:58.080]  so you can go and check it out.
[00:58.080 --> 01:00.640]  But we're going to give you today a brief overview of
[01:00.640 --> 01:03.920]  how to get started and start using it today.
[01:03.920 --> 01:07.600]  Let's talk briefly about why we decided to create it,
[01:07.600 --> 01:10.400]  and there was a need for the project to exist.
[01:10.400 --> 01:13.560]  We could not find a project that had
[01:13.560 --> 01:16.400]  all the variables that LibreTranslate can offer.
[01:16.400 --> 01:19.480]  These are a simple and open-res API that you can
[01:19.480 --> 01:21.960]  use to programmatically do translations,
[01:21.960 --> 01:24.000]  so help automate part of
[01:24.000 --> 01:27.760]  the translation work that we need for the work.
[01:27.760 --> 01:32.600]  It offers pre-trained and openly licensed language models.
[01:32.600 --> 01:35.720]  There are other projects that do machine translation,
[01:35.720 --> 01:37.600]  but again, sometimes they do not make
[01:37.600 --> 01:41.880]  the AI models for the translation openly available,
[01:41.880 --> 01:43.200]  and we have that.
[01:43.200 --> 01:45.920]  Finally, it runs again on commodity hardware,
[01:45.920 --> 01:48.320]  so it does now require server-scale power
[01:48.320 --> 01:50.120]  to make the software work,
[01:50.120 --> 01:53.920]  and finally, it is very easy to get started as you will see.
[01:53.920 --> 01:56.600]  So talking about getting started,
[01:56.600 --> 01:58.600]  there are primarily two ways that you can
[01:58.600 --> 02:00.920]  get LibreTranslate to work on your computer.
[02:00.920 --> 02:03.200]  The first one is if you have Python,
[02:03.200 --> 02:06.600]  you can simply run a pip install command,
[02:06.600 --> 02:11.360]  LibreTranslate, and afterwards you run the program, and that's it.
[02:11.360 --> 02:14.840]  If you have Docker, which many developers like to use,
[02:14.840 --> 02:16.080]  we also have an option for that.
[02:16.080 --> 02:20.120]  We pre-build images for LibreTranslate that you can use,
[02:20.120 --> 02:24.680]  and we have a convenient script that will run it for you and take
[02:24.680 --> 02:27.680]  care of a few details that let you have things like
[02:27.680 --> 02:31.640]  persistent volumes for downloading language models and some technical stuff.
[02:31.640 --> 02:34.800]  But to get started, all you need to do is go on GitHub,
[02:34.800 --> 02:39.240]  get a copy of our source code, and press Run.
[02:39.240 --> 02:43.920]  We also have scripts for Windows and Mac OS and Linux,
[02:43.920 --> 02:46.560]  so we try to support all major platforms.
[02:46.560 --> 02:49.240]  We're hoping to get other platforms in there as well,
[02:49.240 --> 02:53.960]  so things like FreeBSD and others are on the to-do list.
[02:53.960 --> 02:55.680]  So we'll get there.
[02:55.680 --> 02:58.280]  So let's actually try to run it,
[02:58.280 --> 03:01.360]  and I'm always a little scared of doing live demos,
[03:01.360 --> 03:03.920]  but bear with me, we're going to try it.
[03:03.920 --> 03:05.800]  What could go wrong?
[03:05.800 --> 03:08.920]  So here is a console.
[03:08.920 --> 03:11.240]  I'm going to quickly activate
[03:11.240 --> 03:14.760]  a Python environment where I have LibreTranslate already installed,
[03:14.760 --> 03:16.720]  and I'm going to try to run it.
[03:16.720 --> 03:18.640]  On Mac OS, I have to specify
[03:18.640 --> 03:21.840]  a different port than the default of 5,000.
[03:21.840 --> 03:23.200]  I'm going to try to run it.
[03:23.200 --> 03:25.000]  Okay, it seems to be working.
[03:25.000 --> 03:29.400]  So I'm going to jump back right into Chrome,
[03:29.400 --> 03:32.280]  and if I refresh the page,
[03:32.280 --> 03:35.880]  you will be presented with a friendly user interface that you can use
[03:35.880 --> 03:38.920]  to test the system and even use it.
[03:38.920 --> 03:43.800]  It allows programmatic access to the software via an API,
[03:43.800 --> 03:45.280]  but you can also use it as
[03:45.280 --> 03:47.680]  an alternative to Google Translate if you want to.
[03:47.680 --> 03:51.720]  So we're going to try to say something.
[03:51.720 --> 03:55.720]  Okay, so obviously English to English is not going to be helpful.
[03:55.720 --> 03:57.360]  How about French?
[03:57.360 --> 04:00.760]  Okay. So we translated Hello World,
[04:00.760 --> 04:04.880]  Bonjour Le Monde, and it worked.
[04:04.880 --> 04:07.760]  But that's not too impressive, right?
[04:07.760 --> 04:09.320]  I'm like, okay, Hello World.
[04:09.320 --> 04:14.320]  Let's try to look at something a little more realistic.
[04:16.640 --> 04:19.040]  Before looking at something more realistic,
[04:19.040 --> 04:22.960]  you can also, of course, use it from an API.
[04:22.960 --> 04:26.360]  In this case, I can invoke a Cura command and
[04:26.360 --> 04:29.520]  ask Libre Translate to perform a translation.
[04:29.520 --> 04:32.840]  I wanted to automatically detect the language,
[04:32.840 --> 04:34.720]  where the translation is coming from,
[04:34.720 --> 04:38.080]  and finally, I want to translate
[04:38.080 --> 04:42.800]  into the target language.
[04:42.800 --> 04:44.800]  I get a JSON response.
[04:44.800 --> 04:47.240]  Everything in the API is JSON-based.
[04:47.240 --> 04:50.320]  So that will be familiar with many developers.
[04:50.320 --> 04:53.360]  But let's look at a more realistic example.
[04:53.360 --> 04:55.920]  In this case, we have a longer piece of text,
[04:55.920 --> 04:59.040]  and it also contains HTML.
[04:59.040 --> 05:02.080]  The software is capable of translating
[05:02.080 --> 05:03.720]  the parts that need translation while
[05:03.720 --> 05:05.840]  leaving the HTML part intact.
[05:05.840 --> 05:10.160]  So things like hyperlinks do not get mistakenly translated,
[05:10.160 --> 05:13.200]  which would be really bad.
[05:13.200 --> 05:16.760]  This code that we saw here roughly gets
[05:16.760 --> 05:21.280]  represented as this piece of HTML in a browser,
[05:21.280 --> 05:25.360]  and the translation is pretty good.
[05:25.360 --> 05:28.960]  Kind of. This word should have been filidae.
[05:28.960 --> 05:32.880]  It decided to keep the translation in French.
[05:32.880 --> 05:36.080]  We will improve that with time.
[05:36.080 --> 05:38.960]  But otherwise, the context and
[05:38.960 --> 05:41.680]  the meaning of the sentence is pretty darn good.
[05:41.680 --> 05:45.240]  We will look at accuracy in the upcoming slides.
[05:45.240 --> 05:49.320]  So as an overview of the list of features,
[05:49.320 --> 05:50.920]  it can do text translation,
[05:50.920 --> 05:54.120]  it can do markup translation that includes HTML,
[05:54.120 --> 05:57.480]  XML, and other formats that use a markup.
[05:57.480 --> 06:01.520]  It can do several formats for file translation.
[06:01.520 --> 06:03.760]  So you can upload things like
[06:03.760 --> 06:06.000]  open office, LibreOffice,
[06:06.000 --> 06:10.240]  Word documents, and PowerPoint slides,
[06:10.240 --> 06:13.080]  and able to translate those as well.
[06:13.080 --> 06:16.080]  It can perform a language detection.
[06:16.080 --> 06:18.200]  So you give it a piece of text,
[06:18.200 --> 06:20.000]  and it will give you
[06:20.000 --> 06:23.920]  an estimate of which language the program thinks it is.
[06:23.920 --> 06:28.400]  It also has a built-in system for doing rate limiting.
[06:28.400 --> 06:30.840]  If you're planning to host this on a public server,
[06:30.840 --> 06:32.920]  you will find out that it's a very useful feature,
[06:32.920 --> 06:37.400]  because people really like free resources,
[06:37.400 --> 06:42.480]  and it's difficult to give everything for free without
[06:42.480 --> 06:46.840]  some limits. So if your translation instance up in
[06:46.840 --> 06:48.360]  the Cloud gets really popular,
[06:48.360 --> 06:50.840]  having some limit by saying,
[06:50.840 --> 06:53.960]  do a maximum of 60 translation per minute,
[06:53.960 --> 06:56.800]  will come really handy and it's all built in into the software.
[06:56.800 --> 07:02.360]  You can further issue API keys to give to people that can change those limits.
[07:02.360 --> 07:05.280]  So you can set up the system in a way where you
[07:05.280 --> 07:10.560]  allow anonymous users to translate up to 20 translations per minute,
[07:10.560 --> 07:13.800]  and you can allow a subset of people that you've issued
[07:13.800 --> 07:17.160]  API keys to have however many they want.
[07:17.160 --> 07:18.720]  You decide those limits.
[07:18.720 --> 07:21.720]  It also has a localized UI.
[07:21.720 --> 07:23.640]  We're using WebLater to do that,
[07:23.640 --> 07:27.160]  which is awesome, and it has been
[07:27.160 --> 07:29.800]  currently translated into four languages,
[07:29.800 --> 07:32.400]  and we're looking to verify and add more.
[07:32.400 --> 07:34.560]  One cool neat feature is that
[07:34.560 --> 07:40.400]  a Libre translator has a ability to translate itself roughly,
[07:40.400 --> 07:42.920]  so we have done that of course,
[07:42.920 --> 07:48.080]  but we haven't displayed all the languages that it has tried to translate itself.
[07:48.080 --> 07:51.360]  We are waiting for a native speaker to review
[07:51.360 --> 07:53.920]  the actual translation and correct it.
[07:53.920 --> 07:55.520]  But if you run in debug mode,
[07:55.520 --> 07:58.280]  you will see all the work that it has done, which is neat.
[07:58.280 --> 08:02.240]  So it translates itself or at least it helps.
[08:02.240 --> 08:04.640]  It finally has the ability to monitor itself.
[08:04.640 --> 08:07.960]  So it can generate the usage metrics,
[08:07.960 --> 08:13.520]  so you can monitor the usage of the server using Prometheus and Grafana.
[08:13.520 --> 08:17.520]  These are tools to do monitoring that are very popular.
[08:17.520 --> 08:21.920]  Inside the software, there is really just a few packages,
[08:21.920 --> 08:23.680]  so it's very lightweight.
[08:23.680 --> 08:27.360]  Most of the translation work is done by
[08:27.360 --> 08:29.880]  another package called Argos Translate.
[08:29.880 --> 08:32.360]  This is really the core engine that
[08:32.360 --> 08:35.160]  performs the hard work in the translation,
[08:35.160 --> 08:36.920]  which is an awesome project,
[08:36.920 --> 08:40.480]  and we collaborate with them on Libre Translate.
[08:40.480 --> 08:42.280]  Inside Argo Translate,
[08:42.280 --> 08:47.080]  there is also other software which is built on the shoulder of giants.
[08:47.080 --> 08:49.840]  C-Translate, which is an inference engine that
[08:49.840 --> 08:53.200]  does neural translation using transformers models,
[08:53.200 --> 08:54.600]  which is a state-of-the-art.
[08:54.600 --> 08:58.560]  It's the same type of architecture that chatGPT3 uses.
[08:58.560 --> 09:00.440]  There is a sentence piece,
[09:00.440 --> 09:03.200]  which is a piece of code from Google that does
[09:03.200 --> 09:08.040]  the word tokenization and the stanza which comes out of Stanford,
[09:08.040 --> 09:10.480]  which does a sentence analysis.
[09:10.480 --> 09:15.360]  Argos Translate uses all these three to perform the translation work.
[09:15.360 --> 09:17.240]  Now, that's not all it does.
[09:17.240 --> 09:18.960]  Argos Translate also takes care of
[09:18.960 --> 09:23.240]  the very important Argos package manager index.
[09:23.240 --> 09:27.040]  This is where all the language models are handled,
[09:27.040 --> 09:28.520]  installed, and distributed.
[09:28.520 --> 09:31.880]  So the first time that you run Libre Translate,
[09:31.880 --> 09:34.080]  Argos Translate will take care of
[09:34.080 --> 09:36.760]  querying the Argos package manager index,
[09:36.760 --> 09:39.280]  and will download the languages that you need.
[09:39.280 --> 09:42.840]  This allows us to also create instances where,
[09:42.840 --> 09:46.560]  say, you only need to translate between French and English.
[09:46.560 --> 09:51.720]  You do not need to download the entire 26 gigabytes of models.
[09:51.720 --> 09:52.840]  You can simply say,
[09:52.840 --> 09:54.560]  I just need those two models,
[09:54.560 --> 09:58.520]  and the program will download simply those two models.
[09:58.520 --> 10:01.200]  We also have a small module that does
[10:01.200 --> 10:05.320]  the file translation which connects again to Argos Translate.
[10:05.320 --> 10:08.080]  That's the Argos Translate files package,
[10:08.080 --> 10:11.280]  and then some common Python packages that allow us to put
[10:11.280 --> 10:15.560]  the web interface and coordinate the application as a whole.
[10:15.560 --> 10:21.440]  So it's really an ecosystem that's built with other open source software,
[10:21.440 --> 10:27.320]  and together it creates this complete translation solution.
[10:27.320 --> 10:29.320]  Talking about language models,
[10:29.320 --> 10:32.080]  we have 58 of them that gives you
[10:32.080 --> 10:34.560]  translation support for about 30 languages.
[10:34.560 --> 10:38.720]  It does automatic pivot via English.
[10:38.720 --> 10:43.440]  We are currently looking to transition to using multi-language models.
[10:43.440 --> 10:45.800]  But for the moment when you translate,
[10:45.800 --> 10:47.640]  say, from Italian to French,
[10:47.640 --> 10:51.200]  the program will automatically do the pivoting via English.
[10:51.200 --> 10:55.360]  So I will translate Italian to English and English to French.
[10:55.360 --> 10:57.600]  If there is a language missing,
[10:57.600 --> 11:01.560]  there is a very cool repository
[11:01.560 --> 11:04.520]  under the Argos OpenTech organization,
[11:04.520 --> 11:07.840]  which builds Argos Translate called Argos Train.
[11:07.840 --> 11:09.800]  That is a repository that has
[11:09.800 --> 11:13.520]  very good instructions on how you can train your own models.
[11:13.520 --> 11:17.520]  So if a language is missing, go check it out.
[11:17.520 --> 11:20.280]  It has very clear instructions,
[11:20.280 --> 11:22.880]  and you could contribute a language that is
[11:22.880 --> 11:26.480]  missing and you want to see integrated into the software.
[11:26.480 --> 11:29.920]  Speaking of the models,
[11:29.920 --> 11:32.040]  when a model is downloaded,
[11:32.040 --> 11:34.440]  it has a Argos model extension,
[11:34.440 --> 11:36.880]  and these are simply zip files.
[11:36.880 --> 11:39.440]  It's a zip file and it's inside,
[11:39.440 --> 11:42.440]  has a little bit of metadata.
[11:42.440 --> 11:46.480]  It has a folder that contains the CTranslate model.
[11:46.480 --> 11:49.960]  It has the sentence piece model and finally the stanza model.
[11:49.960 --> 11:52.560]  So it has the information for all the three packages that we
[11:52.560 --> 11:55.200]  discussed earlier to perform the translation.
[11:55.200 --> 11:58.000]  It's very interesting to check it out.
[11:58.000 --> 12:00.360]  Let's talk a little bit of accuracy, right?
[12:00.360 --> 12:01.880]  Like the question like, okay,
[12:01.880 --> 12:05.320]  it's a translation sorry, but how good is it really?
[12:05.320 --> 12:09.560]  For that, there is a metric that can be used to
[12:09.560 --> 12:13.800]  assess roughly the accuracy of the translation.
[12:13.800 --> 12:15.920]  It's called a Blue Score acronym
[12:15.920 --> 12:18.720]  for bilingual evaluation under study.
[12:18.720 --> 12:25.160]  It measures the similarity of text to a reference corpus.
[12:25.160 --> 12:28.160]  It has values that go from 0 to 1,
[12:28.160 --> 12:31.960]  or if you express it as a percentage from 0 to 100.
[12:31.960 --> 12:34.240]  The best translators in the world,
[12:34.240 --> 12:37.840]  human translators do not get a score of 100 ever.
[12:37.840 --> 12:41.400]  So anything that is above
[12:41.400 --> 12:45.040]  a 40 is considered understandable to good.
[12:45.040 --> 12:49.720]  Something that is above 50 tends to be very high quality.
[12:49.720 --> 12:55.120]  Sorry, up to 50 is high quality and above 60 is very high.
[12:55.120 --> 12:58.240]  We had a community contributor
[12:58.240 --> 13:00.840]  actually go and a few weeks ago,
[13:00.840 --> 13:06.080]  he ran the evaluation on our different models,
[13:06.080 --> 13:08.840]  and we found that 83 percent of
[13:08.840 --> 13:10.480]  the models currently in
[13:10.480 --> 13:13.720]  Libre Translate are scoring above 40 percent.
[13:13.720 --> 13:16.880]  So 83 of them are good.
[13:16.880 --> 13:19.840]  Now, to make it into perspective,
[13:19.840 --> 13:23.280]  when people ask me directly how good is Libre Translate,
[13:23.280 --> 13:26.000]  I like to tell them that it's roughly as good as
[13:26.000 --> 13:28.600]  Google Translate was four years ago.
[13:28.600 --> 13:31.440]  So I want to make the expectations clear at this stage in
[13:31.440 --> 13:33.640]  the project that it is not as good as some of
[13:33.640 --> 13:36.000]  the proprietary alternatives.
[13:36.000 --> 13:40.400]  But we are improving and we will continue to improve.
[13:41.600 --> 13:47.320]  The way to improve it lies into mostly getting better training data.
[13:47.320 --> 13:50.200]  So as we find more and more sources of
[13:50.200 --> 13:52.480]  open data that can be used for translation,
[13:52.480 --> 13:54.560]  we include those into the training of
[13:54.560 --> 13:58.480]  the models and that results into better models.
[13:58.480 --> 14:03.080]  This is also an interesting point to note,
[14:03.080 --> 14:08.560]  is that because the project is open source and we have a way to train models,
[14:08.560 --> 14:12.440]  you can also train models that are specific to a certain domain.
[14:12.440 --> 14:15.040]  For example, in the context of software translation,
[14:15.040 --> 14:18.960]  you could imagine the case where instead of training the data on
[14:18.960 --> 14:27.480]  a general corpus like Wikipedia or the EU Parliament translation documents,
[14:27.480 --> 14:31.320]  you could train a model that is specific to software.
[14:31.320 --> 14:36.400]  For example, you could take a set of existing translations from
[14:36.400 --> 14:39.320]  existing software that has licensed
[14:39.320 --> 14:43.760]  the translation work under an open permissible license and train
[14:43.760 --> 14:47.240]  a model onto those existing translations.
[14:47.240 --> 14:49.640]  Because we have the knowledge,
[14:49.640 --> 14:51.960]  a lot of software has commonalities in terms.
[14:51.960 --> 14:53.720]  When you have a file menu,
[14:53.720 --> 14:56.280]  it's always called file and then edit.
[14:56.280 --> 15:00.440]  So those menus are specific to a context.
[15:00.440 --> 15:04.120]  By training models that are specific to a context,
[15:04.120 --> 15:06.960]  you could get, for example,
[15:06.960 --> 15:12.200]  software translation model that is more accurate in the context of software,
[15:12.200 --> 15:13.840]  rather than say poetry.
[15:13.840 --> 15:18.880]  So it's a very interesting thing to think about.
[15:18.880 --> 15:21.000]  One more thing about accuracy,
[15:21.000 --> 15:23.560]  we do have the occasional rare quirk.
[15:23.560 --> 15:27.400]  This is something that we are aware of and we are working to fix it.
[15:27.400 --> 15:31.400]  We like to call it the salad issue.
[15:31.400 --> 15:38.240]  We joke, I will demonstrate this slide because it always sparks a little bit of a giggle.
[15:40.800 --> 15:43.360]  It's a little bit rare, but it happens.
[15:43.360 --> 15:46.880]  So in Spanish, the word for salad is ensalada.
[15:46.880 --> 15:51.520]  Now, let's try to translate the word for salads plural.
[15:51.520 --> 15:55.080]  So I'm going to type ensaladas.
[15:55.080 --> 15:58.360]  So in French, that's saladas.
[15:58.360 --> 16:00.080]  Is that correct? Any French people in the room?
[16:00.080 --> 16:02.320]  Fantastic.
[16:02.320 --> 16:05.320]  Now let's try the singular form.
[16:05.320 --> 16:10.040]  I'm going to remove the S and it crunches for a little bit.
[16:10.040 --> 16:18.840]  In a second, it really likes salad.
[16:18.840 --> 16:21.840]  Salad, salad, salad, salad, salad.
[16:21.840 --> 16:24.560]  This is a quirk. We are aware of it.
[16:24.560 --> 16:29.640]  It's very rare, but we've found a few reports here and there and we're working to fix it.
[16:29.640 --> 16:31.960]  Just something to be aware of.
[16:33.160 --> 16:35.640]  Yes, it really likes salad.
[16:35.640 --> 16:40.000]  Me too. Let's talk a little bit about integrations.
[16:40.000 --> 16:47.360]  You can find the client libraries for about 11 programming languages that includes the most common ones like Java, Python,
[16:47.360 --> 16:52.400]  whatever your favorite language is, it's probably in the list of bindings.
[16:52.400 --> 16:57.200]  And if it's not there, adding new bindings for LibreTranslate is fairly easy.
[16:57.200 --> 17:00.000]  So we welcome contributions, of course.
[17:00.000 --> 17:07.000]  As far as software, LibreTranslate has found adoption in several existing open-source software that you may recognize.
[17:07.000 --> 17:11.760]  Mastodon recently added support for translating topics using LibreTranslate.
[17:11.760 --> 17:21.160]  Weblate has the ability to use LibreTranslate to suggest and help translators perform translations as an alternative to using proprietary software.
[17:21.160 --> 17:30.720]  The forum software discourse has a plugin that lets you make your forum software accessible from different locales
[17:30.720 --> 17:34.520]  and lets you translate the posts on the fly.
[17:34.520 --> 17:40.960]  LibreOffice, I found, has an extension. I didn't know this until a week ago when I was looking who has integrated stuff with LibreTranslate.
[17:40.960 --> 17:45.960]  But somebody wrote an extension to LibreOffice where you can translate documents on the fly using LibreTranslate.
[17:45.960 --> 17:49.360]  There is an add-on for the multimedia software code.
[17:49.360 --> 17:51.960]  There is an add-on also for Firefox.
[17:51.960 --> 17:55.560]  And there's probably a lot of other things that I haven't found myself.
[17:55.560 --> 18:01.760]  But a lot of people seem to be finding the API useful and they're doing integration work, which is fantastic.
[18:01.760 --> 18:08.320]  And there's finally client applications that you can use LibreTranslate with without using the web UI.
[18:08.320 --> 18:12.320]  And we found we have clients for Android, iOS and desktop.
[18:12.320 --> 18:17.320]  And there's more being built by the week.
[18:17.320 --> 18:25.760]  As far as comparison to proprietary alternatives, you can see that there is a clear monetary advantage,
[18:25.760 --> 18:32.160]  aside from the philosophical reason for why you might want to use open-source software, of course.
[18:32.160 --> 18:40.160]  But it could also be a really sustainable way to perform translations in that people often ask me,
[18:40.160 --> 18:43.360]  why should I use LibreTranslate? I can use Google Translate for free.
[18:43.360 --> 18:48.160]  I just go on translate.google.com and it doesn't charge me anything. So why should I care?
[18:48.160 --> 18:52.160]  Google Translate is free so long as you're using it by hand.
[18:52.160 --> 18:57.760]  If you want to do any automation work and you have to tap into their API, you're going to pay dearly.
[18:57.760 --> 19:03.960]  And you can see here a list of the prices and I can assure you that one million characters seem like a lot,
[19:03.960 --> 19:12.960]  that's six zeros, but they actually run pretty fast and so could the bill on your credit card.
[19:12.960 --> 19:20.760]  So if you have a lot of text to translate, LibreTranslate could really help in that regard.
[19:20.760 --> 19:26.560]  As far as funding goes, the project is on the path to become fully self-funded.
[19:26.560 --> 19:31.360]  And we really care about this because we want the project to continue living on.
[19:31.360 --> 19:36.160]  We, of course, accept sponsorships and donations, but honestly,
[19:36.160 --> 19:41.360]  we would rather prefer that you get something back if you decide to contribute financially to the project.
[19:41.360 --> 19:48.360]  This is why if you are in the position where you say, I have some finances to spare and help support the project,
[19:48.360 --> 19:58.160]  you also get something back and we do that in the form of offering you an API key to use a host distance at LibreTranslate.com.
[19:58.160 --> 20:04.960]  So you are free to run the infrastructure on your own server, on your Raspberry Pi, on any machine that you'd like.
[20:04.960 --> 20:12.360]  If you don't want to handle that, you can just get an API key and you can support the project at the same time.
[20:12.360 --> 20:15.560]  So it's really a good way to contribute back.
[20:15.560 --> 20:20.960]  And we found that that model has been helping us grow and sustain the project.
[20:20.960 --> 20:26.160]  So we hope to continue growing as much next year.
[20:26.160 --> 20:29.960]  Again, to get involved, I'll give you a few quick numbers.
[20:29.960 --> 20:34.160]  We've had about 70 people contribute to the code base over the last few years.
[20:34.160 --> 20:40.960]  The project is still very young, but it has really received a lot of attention, so we're very excited about that.
[20:40.960 --> 20:48.760]  You can help with code. If you're a Python programmer, if you know HTML, CSS, any of the technologies that we use, you're welcome to contribute.
[20:48.760 --> 20:51.860]  We are open to everybody and all ideas.
[20:51.860 --> 20:54.560]  You can also help us translate.
[20:54.560 --> 21:05.160]  If you understand English and you don't see your language in the list of languages that we currently support for your user interface, you are welcome to contribute.
[21:05.160 --> 21:12.360]  It's on a web late, you can simply translate and it will get included into the project every 24 hours.
[21:12.360 --> 21:14.960]  So that is really amazing.
[21:14.960 --> 21:18.160]  You can also help us train more language models.
[21:18.160 --> 21:28.460]  If your language is not available or a language that you care about is not available, you can yourself create a new model for a language and add that into the list.
[21:28.460 --> 21:31.060]  So that is also another way that people can help.
[21:31.060 --> 21:37.560]  You can report bugs, of course. If you don't report salad, we are aware of it.
[21:37.560 --> 21:38.760]  Or just come say hi.
[21:38.760 --> 21:46.360]  We have a community forum that is quickly growing and we love to hear what you're building with it, what you're using, or if you have any questions.
[21:46.360 --> 21:51.760]  So we're very, very open and we're excited to hear what you will do with it.
[21:51.760 --> 21:54.360]  That said, this was the last slide.
[21:54.360 --> 21:57.160]  I think we have some time left over, right?
[21:57.160 --> 22:02.160]  So I will.
[22:02.160 --> 22:03.760]  So thank you very much.
[22:03.760 --> 22:05.960]  I will open the floor for questions and discussion.
[22:05.960 --> 22:08.360]  So yes.
[22:08.360 --> 22:20.760]  Hi, my best friend cannot miss the Vice President of the Austrian Society for Artificial Intelligence sitting with Foster to find volunteers to do exactly what you're doing.
[22:20.760 --> 22:27.760]  Thank you. We're glad to be able to help.
[22:27.760 --> 22:28.760]  You're welcome.
[22:28.760 --> 22:34.560]  How do we find, well, I just named the thing Open Language Model Training Army.
[22:34.560 --> 22:36.960]  How do we find more volunteers?
[22:36.960 --> 22:47.260]  Unemployed people, maybe have the government fund people running training models, maybe suggest that to all politicians, everybody to their member of parliament.
[22:47.260 --> 22:51.660]  How many people do we have here?
[22:51.660 --> 22:55.060]  Should be all of Europe at least, maybe South America.
[22:55.060 --> 22:58.760]  I think if we multiply this, it can go viral.
[22:58.760 --> 22:59.960]  Thank you very much.
[22:59.960 --> 23:01.060]  This is awesome work.
[23:01.060 --> 23:01.360]  Thank you.
[23:01.360 --> 23:03.360]  I appreciate it.
[23:03.360 --> 23:10.260]  Yes, but you speak of our language that have the same link, the same structure of the language.
[23:10.260 --> 23:15.460]  We have French, English, Spanish, Portuguese, maybe Russian and Ukrainian.
[23:15.460 --> 23:20.560]  I do not have the same structure, but the language not far away from here.
[23:20.560 --> 23:22.460]  That's the difference.
[23:22.460 --> 23:24.660]  German also.
[23:24.660 --> 23:26.160]  Correct.
[23:26.160 --> 23:32.560]  And so taking this in account, it's not easy for a translator, for them to translate it.
[23:32.560 --> 23:33.160]  It is not.
[23:33.160 --> 23:38.360]  There's also maybe a problem with Chinese or Japanese.
[23:38.360 --> 23:39.260]  Correct.
[23:39.260 --> 23:50.760]  There was a problem, there was a thing I went to say, it's a dictionary in line or in the program to have the good word because it's not translated every time the good word.
[23:50.760 --> 23:57.160]  And so I thought also the most efficient people's language is Esperanto, not English.
[23:57.160 --> 23:58.860]  Oh, that is very interesting.
[23:58.860 --> 23:59.460]  Yes.
[23:59.460 --> 24:00.160]  Okay.
[24:00.160 --> 24:02.360]  Yeah, that's a great insight.
[24:02.360 --> 24:04.760]  Yeah, thank you for sharing that.
[24:04.760 --> 24:13.560]  And you're completely right, some languages don't share the same semantical structure and Dutch, for example, currently doesn't score super high.
[24:13.560 --> 24:18.160]  It's actually one of the bottom 17% of the language models in the blue score.
[24:18.160 --> 24:20.660]  Dutch scored around 38%.
[24:20.660 --> 24:28.860]  So it's almost good, but we've had some Dutch speaking people come to us and say, you know, it's like equal use improvement.
[24:28.860 --> 24:31.460]  So Dutch, yes, it is a language that needs improvement.
[24:31.460 --> 24:40.060]  And I talked to the maintainer of Argus translate about the languages that need improvement.
[24:40.060 --> 24:46.360]  And he pretty much suggested that better training data will help greatly.
[24:46.360 --> 24:51.660]  So it is mainly a problem, not of the architecture of the AI.
[24:51.660 --> 25:01.760]  It's a matter that we don't have sufficient quality, high quality data between, say, English and Dutch to get above 38% currently.
[25:01.760 --> 25:04.960]  But again, nobody has really focused on Dutch as a language.
[25:04.960 --> 25:10.860]  If anybody has an interest in improving Dutch, we can do better.
[25:10.860 --> 25:16.460]  Surprisingly, fantastic.
[25:16.460 --> 25:21.960]  But as far as, for example, languages like German, LibreTranslate currently does very well with German.
[25:21.960 --> 25:24.460]  It's above 50, if I remember correctly.
[25:24.460 --> 25:27.360]  Is it because German is the similar language to Dutch?
[25:27.360 --> 25:28.360]  It is.
[25:28.360 --> 25:35.560]  That is because I believe, and I think PJ, that's the name of the maintainer of Argus translate,
[25:35.560 --> 25:45.460]  because the German model has had a larger amount of training data, and so it tends to perform better.
[25:45.460 --> 25:46.460]  Yes?
[25:46.460 --> 25:51.460]  Yeah, just a quick question around the translation process, I suppose, touched on the structure.
[25:51.460 --> 25:54.460]  But how does it work with different dialects?
[25:54.460 --> 25:57.460]  So if you write in dialects, will you write in slang?
[25:57.460 --> 25:58.960]  That's a very good question.
[25:58.960 --> 26:05.460]  Dialects would probably, and that's my guess, because I've never inquired this myself,
[26:05.460 --> 26:12.460]  but I believe that a dialect to perform good as a target or source language for translation
[26:12.460 --> 26:16.460]  would also need its fair amount of training data.
[26:16.460 --> 26:18.460]  And that is the problem with dialects.
[26:18.460 --> 26:22.460]  I actually speak a local Italian dialect, that is my first language,
[26:22.460 --> 26:26.460]  and I wanted to make a model for my dialect.
[26:26.460 --> 26:31.460]  And I started looking online for references of data that I could use to create a model for my dialect,
[26:31.460 --> 26:33.460]  because it would be cool.
[26:33.460 --> 26:36.460]  And it was really challenging.
[26:36.460 --> 26:40.460]  Not being an official language, it really lacks the status of official languages,
[26:40.460 --> 26:44.460]  and finding training data is extremely difficult.
[26:44.460 --> 26:46.460]  But it could be possible, right?
[26:46.460 --> 26:54.460]  If you gather enough people that can create a ground truth data set of examples in the dialect
[26:54.460 --> 26:58.460]  with sufficient samples, you could get good results, I believe.
[26:58.460 --> 27:02.460]  So it's a matter, again, of training data.
[27:02.460 --> 27:03.460]  Yes?
[27:03.460 --> 27:06.460]  How much does it cost to get a model to a good level?
[27:06.460 --> 27:12.460]  In terms of computing power or in terms of computing power?
[27:12.460 --> 27:19.460]  So if I remember correctly what PJ told me about the cost of training the models,
[27:19.460 --> 27:26.460]  it costs maybe a few, between $12 and $30.
[27:26.460 --> 27:29.460]  You can rent instances on several cloud providers.
[27:29.460 --> 27:31.460]  You do need a GPU to train these models,
[27:31.460 --> 27:37.460]  and it might take a few days for it to crunch and get sufficient number of iterations to train the model.
[27:37.460 --> 27:40.460]  But it's absolutely affordable.
[27:40.460 --> 27:45.460]  Anybody can do it, and if you are willing to wait and you just have a gaming laptop sitting at home,
[27:45.460 --> 27:50.460]  if you're OK waiting 20 days for it to finish, it will train the model for you.
[27:50.460 --> 27:55.460]  So I guess it could be free to you if you're willing to wait a sufficient amount of time,
[27:55.460 --> 27:59.460]  and if you have a gaming laptop lying around.
[27:59.460 --> 28:00.460]  Yes?
[28:00.460 --> 28:06.460]  So you mentioned the need for a data availability for doing the model, right?
[28:06.460 --> 28:13.460]  Well, do you need the data to be available under a certain license?
[28:13.460 --> 28:15.460]  What's your problem?
[28:15.460 --> 28:19.460]  The world's full of things, right?
[28:19.460 --> 28:20.460]  Yes.
[28:20.460 --> 28:21.460]  What's the requirement you have?
[28:21.460 --> 28:24.460]  You want it to be public domain?
[28:24.460 --> 28:34.460]  It has to be licensed under a permissive license, so creative comments that also includes commercial use.
[28:34.460 --> 28:42.460]  And we give references and we give attribution to all the sources that we use.
[28:42.460 --> 28:50.460]  If you go into the Argos Package Manager repository, where all the models are hosted,
[28:50.460 --> 28:53.460]  we do give the appropriate licensing credits to all those.
[28:53.460 --> 28:58.460]  But yes, we cannot go on, say, the Internet and start scraping results,
[28:58.460 --> 29:04.460]  because everything, you just have to assume that everything is covered by copyright
[29:04.460 --> 29:06.460]  until they tell you that you can use it freely.
[29:06.460 --> 29:12.460]  So it's only trained on openly available and freely licensed sources.
[29:12.460 --> 29:17.460]  Do you need it to be translated as well or just a single language?
[29:17.460 --> 29:19.460]  It has to be translated.
[29:19.460 --> 29:27.460]  So very briefly, the format of the input that goes into the training is a file that has,
[29:27.460 --> 29:33.460]  say, the English sentences and a separate file that has the translation on the same line.
[29:33.460 --> 29:36.460]  So it's very basic.
[29:36.460 --> 29:39.460]  And somebody could do the work by hand, right?
[29:39.460 --> 29:43.460]  You start from the English translation and you start doing the translation.
[29:43.460 --> 29:49.460]  So it will take a lot of work, but it's doable, especially in a crowd-formed...
[29:49.460 --> 29:51.460]  Are we out of time?
[29:51.460 --> 29:52.460]  Okay.
[29:52.460 --> 29:53.460]  I'll be around if you have other questions.
[29:53.460 --> 29:55.460]  Our time is up, unfortunately.
[29:55.460 --> 29:56.460]  They're kicking me out.
[29:56.460 --> 30:00.460]  But the next speaker will deliver something awesome as well next talk.
[30:00.460 --> 30:15.460]  Thank you again.