Yandex has decided to astonish everyone and has made the neural network YaLM 100B openly accessible. It allows generating and processing texts in both Russian and English languages. Furthermore, it comprises 100 billion parameters – yes, precisely 100 billion. This is the largest GPT-like model available to the public. According to its creators, the neural network is capable not only of generating product descriptions and video scripts, but can even provide ideas for advertising content.
What is YaLM 100B?
Yandex employs the neural network in 20 projects: Alice, quick answers in search, customer support, generating descriptions in snippets, and more. During the training process, it was fed with around 2 TB of text (if one recalls that a fairly voluminous book fits into 1.5 megabytes, the scale of the work done becomes clear) over 65 days. Yandex's own supercomputers were utilized for this purpose.
If you want to teach YaLM 100B something new, get ready to rent a mining farm: this neural network requires 200 GB of video memory. Yes, that's right – the computations are done on graphics card processors, so you'll need to either invest in 35 Nvidia 1060 cards (which we used to test the neural network's skills) or a similar quantity of other cards.
How to install?
Prepare to free up about 500 GB of available space. The description on GitHub states that the neural network specifically requires 200 GB. However, after the installation process begins, the folder size increases to over 350 GB (and that's not the limit).
What do you need to do to install YaLM 100B (providing one of the installation methods)?
What do you need to do to install YaLM 100B (providing one of the installation methods)?
- Download and install the standalone client. In the settings (if you're not familiar), you can leave everything unchanged.
- Launch Git Bash (you'll be prompted to choose this at the end of the installation process).
- In the opened window, specify where to download the neural network. The commands will be entered as if in cmd or PowerShell windows. In other words, to change the directory, type cd E: (to change the drive) and cd/FolderName (all in English, of course). For convenience, you can type the first letters of the folder name and press Tab (the name will be completed automatically).
- Once you've specified the desired folder, enter the following commands: git lfs install and git clone https://huggingface.co/yandex/yalm-100b --depth 1.
5.Now, allocate some time, free up space on your drive, and wait. The waiting period will be quite long.
After this, the Git Bash client will download and install all necessary files for you. Then, you can begin testing YaLM 100B according to your needs.
IMPORTANT: While the installation is ongoing, avoid pressing the Ctrl+c combination in the Git Bash window if you want to copy the progress status. This will cancel the previous command and the installation (and you'll have to start everything over). To copy, highlight the desired lines with your mouse and right-click.
You'll also need to install:
After installing Python, you can add the necessary modules directly in Git Bash using a simple command like pip3 install deepspeed (or replace "deepspeed" with the relevant module's name at the end of the command).
Once you're prepared, you can move on to testing. The simplest way is to enter the command bash examples/generate_interactive.sh from the installed YaLM directory. This will allow you to interact with the neural network directly from the command line.
After this, the Git Bash client will download and install all necessary files for you. Then, you can begin testing YaLM 100B according to your needs.
IMPORTANT: While the installation is ongoing, avoid pressing the Ctrl+c combination in the Git Bash window if you want to copy the progress status. This will cancel the previous command and the installation (and you'll have to start everything over). To copy, highlight the desired lines with your mouse and right-click.
You'll also need to install:
- Python (with modules like torch, deepspeed, six, etc.).
- CUDA (for offloading computations to the graphics card processor).
After installing Python, you can add the necessary modules directly in Git Bash using a simple command like pip3 install deepspeed (or replace "deepspeed" with the relevant module's name at the end of the command).
Once you're prepared, you can move on to testing. The simplest way is to enter the command bash examples/generate_interactive.sh from the installed YaLM directory. This will allow you to interact with the neural network directly from the command line.
Usage Examples
Let's note right away: as of now, we haven't managed to obtain large unique texts for the offers that interest us. It's possible that additional training is required (and renting a mining farm for this purpose) or refining the queries to the neural network. However, we tried using YaLM 100B to create text for small advertisements.
What did we come up with?
Advertisement for an imaginary weight loss program.
What did we come up with?
Advertisement for an imaginary weight loss program.
Have you ever wanted to lose weight but didn't know where to start? MagicSlim is here for you! This program will teach you how to eat right and stay slim without diets or workouts. Just follow the instructions, and the weight will melt away. Try it today!
Lose Weight for Summer! Find out how to slim down before the start of summer. Begin the course now!
This summer, you'll achieve the figure you've always dreamed of with the premium weight loss course from Manona+.
And we also managed to get a bit of fun trash for teasers. In these examples, we only inserted the first 2-3 words, and the teaser was completed automatically. We will abbreviate or remove celebrity names from the final version for ethical reasons (but you know them all).
A.P. urgently fled Israel after a terrorist attack during the concert! The star has lost...
M.G. has lost face! The singer was humiliated in his own home!
Get pumped with the KGB method! Secrets of Soviet soldiers!
From the editorial team
Currently, YaLM 100B is an "endeavor" for those well-versed in programming (or willing to dedicate ample time), as launching the neural network is best done on Linux, primarily using the Git Bash command line interface – meaning without any external interface.
During the setup and testing phase, we encountered regular crashes, the need to reinstall certain files or libraries, and so forth. For instance, we often faced issues with losing the Deepspeed module.
During the setup and testing phase, we encountered regular crashes, the need to reinstall certain files or libraries, and so forth. For instance, we often faced issues with losing the Deepspeed module.