Desert Ant CLI

[00:00] One of the great things about primarily working with local AI is you get to experiment with a lot of small models trained to do very specific tasks. I use a very specific and small model for translation. I’ve trained a small model to do chaptering of my YouTube videos. And in this video, we’re taking a look at Desert Ant Labs, which is an organization somewhere out of Europe, I’m not sure where, and they seem to be completely focused on creating these exact types of models I’m talking about.

[00:28] So each of these cards here on their website represents a model that is trained to do a very specific task. And most of these models come in at around a hundred megabytes or less. Now they have model repositories for each and every one of these on Hugging Face, and they have a TypeScript SDK, and they have a Python SDK, and a bunch of different ways you can interact with these. But what we’re gonna do is take a look at the Desert Ant CLI, which is a tool we can use on our own machine with very few resources to download and manage all these models and interact with each one according to its specific task.

[01:02] Okay, we can install this guy with brew, install, Desert Ant Labs, tap, Desert Ant. And once that’s set up, we can run it with Desert Ant, or we can run it with just DA, which is my preference, and that’s what I’m gonna use going forward. Now we can run DA cache and see that we currently have no models installed. So let’s take a look at our first model called Voz, which is a transcription model. Now I’ve got a video here, HFCLI. It is about 13 minutes long, give or take, and we’re gonna try to create a transcription with DA, Voz.

[01:38] And then we’re just gonna pass in the video file and the type of transcription which we want, and I’m gonna say SRT. And now the first time you run one of these, the model needs to download. And I’m not even gonna fast forward this or anything because they are super small and this should be really quick. So once it gets done downloading, which it is about to do, it will then begin to load the model into memory. And then once it’s loaded into memory, it’ll transcribe quickly. And here we go. We have our transcription right here, HFCLI. And if I cat that out, we can see we have it.

[02:13] Now just for clarification, the way I would normally do this is I would use FFmpeg to convert this video to a very specific type of wave. Let me drop that in here. And this part’s really quick. Now the next thing I’m gonna do is use Whisper C++, which is my normal process. And I’m gonna say create an SRT. And this part I am gonna fast forward a little bit because you can see it’s taking much, much longer. Okay, that took about 33 seconds, give or take.

[02:45] Whereas with Desert Ant, what we could do is take a look in this batch directory. I have three videos, including a copy of the one that we just processed. And what I’m gonna do is I’m gonna say DAVoz and I’m gonna pass in the folder. So batch forward slash, and I’m gonna say SRT. And we’re gonna watch it do all three faster and then it took Whisper C++ to do the one. So that is one done. And that’s two done. And that’s three videos all transcribed. We’re gonna jump into that folder and list it out. And there we go, we have an SRT for each and every video file.

[03:26] Okay, the next thing we’re gonna take a look at is the clips model. Now what the clips model does is take your video and go through the transcription of it, which it’ll create with Voz. And I’ll double check on the description, but it’s trying to find meaningful clips that you could create from a larger video. So we can see it’s transcribing and now it’s loading clips. Cool, it says it’s done. Let’s take a look. We have 13 clips and I’ve gone through these. They’re very reasonable. Some of them start and stop at kind of awkward moments, but I’m very impressed.

[04:00] All right, for this next one, I need to create some bad audio. So what I’m gonna do is I’m gonna switch from this mic to a much crappier mic. I’m gonna turn on this fan and I’m gonna say, “Um” a bunch of times so that we can take a look at another model. All right, we are looking at Desert and Labs and they’re a small, awesome models. All right, sorry about that. Let’s take a look at what we can do with it. Okay, I extracted the audio from that video and cut it down a little bit and put it in this folder called noise. And there’s our noise MP3.

[04:42] And now what we’re gonna do is we’re gonna run DA clear and our MP3. It’s gonna download the model and it claims it cleared it up. So let’s check it out. So here’s our original noisy file. - Um, Desert and Labs and they’re small, awesome. - And here’s our new file. Um, Desert and Labs and they’re small, awesome. - It is very much improved. I’m impressed. So now I purposely left a bunch of ums and uhs in that recording so that we can take a look at the Uhm model.

[05:25] So what we’re gonna do is we’re gonna run DA uhm and our cleaned up M4A file now. And it’s gonna download the um model. And what it’s going to do is it’s gonna identify everywhere that I said um. And you can see here it’s got the time markers as well as how long the um lasted. So it doesn’t actually remove those moments from the audio but we get those markers and the duration. So we could go in and extract those from the audio ourselves or possibly pass it on to another model to do that for us. So let’s take a look at a couple more of these.

[06:01] The first one we’re gonna check out is Title. Now I have a markdown file here in this text folder. It’s called 100MD. It is an article I wrote about not caring about 100Xing my workflow. I’m gonna cat that and then pipe it to a, uh, to DA Title. And it’s gonna download the title model and it’s gonna come back with what it thinks is a good title for that article. So it came back with AI as a tool for productivity, which is much nicer than the title that I had, uh, and perfectly acceptable.

[06:35] Now another one we can take a look at is GIST and it does more or less exactly what it sounds like. It reviews a piece of content and attempts to give you some categories for it. So we run that and we’ve got technology and software being the strongest and then career and productivity following that perfectly acceptable. The last one we’re gonna take a look at is, uh, called moderator. And what I have here is a directory called images. And in that I have a handful of images. Now I have a website that automatically shares automatic prompt generated images. And every once in a while it produces some not safe for work images. And I have to go in and clean those up or delete them or whatever.

[07:15] Now of these files, uh, two of them are dirty pictures. Uh, they are the first one that begins with four zeros and the last one that begins with four nines. Moderator is supposed to be able to identify which ones are dirty. Pictures are not safe for work. So I’m going to run da moderator and I’m just going to batch the entire images directory. And I’m going to ask for that output in JSON. It’s going to download the model. And here is our output. So the first one was true. The second was false. And the last one here is true. So both of our dirty pictures were correctly identified.

[07:54] So the last thing I want to show you is how to manage these models. And what I’m going to do is I’m going to go ahead and remove them all. Since they’re so small, I could just download them again. And I’m going to run da cache. And we can see we have all these models and the total size of them is one gigabyte. Now, if you want to remove these models, you run da cache double tack clean, but you also need to throw on a double tack force. Otherwise it’ll just tell you, you need to. And now all of my models are gone.

[08:20] There’s a bunch more models available right now, including a handful that are in beta and Desert Ant. If you ever see this video and you want to shoot me access to those beta models, I’d be happy to check them out. But go out and check out Desert Ant Labs. They have some really interesting models that you can run on device or on your machine to do very specific tasks.