Skip to content

Training the first Cypriot Large Language Model on the Masked Language Modeling objective for predicting a given masked word token within a given context

Notifications You must be signed in to change notification settings

pedroandreou/Cypriot-LLM

Repository files navigation

Cypriot LLM

A Cypriot edition of Google's BERT pre-trained language model.



HuggingFace Repo: https://huggingface.co/petros/bert-base-cypriot-uncased-v1

🏗️ Environment

You should create a virtualenv with the required dependencies by running

python -m venv .venv

How to activate the virtual environment to run the code

## Linux
source ./.venv/bin/activate
pip install -r requirements.txt


## Windows
source ./.venv/Scripts/activate
pip install -r requirements.txt

Make a copy of the example environment variables file (this is not the virtual env; don't get confused)

## Linux
cp .env.example .env


## Windows
xcopy .env.example .env

How to update the requirements

When a new requirement is needed you should add it to unpinned_requirements.txt and run

cmd.exe /C setup_new_environment.bat

💻 Cyclone

Cyclone is the name of the super computer at the Cyprus Institute. In order to run a job you need to request access by following this link. Through the process of getting access, you need to pass over your public key so that you can ssh into the super computer.

Each user is being given a separate user directory. In the home directory there is a project folder with the name data_p156, where p156 is the project id. This folder is shared among all members of the project. The project folder also has generous quota for saving data and models so all our work should go there.

If the github repo is not there already, clone and create a virtualenv but to do this, you need to follow below.

Cloning GitHub repo on the supercomputer

Before cloning the github repo, you need to do:

  1. cd ~/.ssh on the supercomputer to copy the content of the id_rsa.pub file which is your public key
  2. Go to your GitHub's Settings - SSH and GPG keys,
  3. Click New SSH key
  4. Add a descriptive title and paste the key into the "Key" field and click Add SSH key.
  5. Try to clone your github repo in the SSH way; for example, by doing git clone [email protected]:pedroandreou/Cypriot-LLM.git

Transferring the raw dataset to the supercomputer

Transfer the dataset from your local machine to the cloned git repo on the supercomputer by downloading WinSCP for Windows. If your local OS is different, then follow this documentation page for doing so.

Cypriot-LLM/
├── dataset/ # transfer it here
├── README.md
└── ...

Modify environment variables file

Go to the .env file and add the path of the dataset as follows:

DATASET_DIR_PATH="/nvme/h/cy22pa1/data_p156/Cypriot-LLM/dataset"

If the repo paths do not work:

Go to the root directory of the git cloned repo and run:

pip install -e .

Login to your HuggingFace account

Login by doing:

huggingface-cli login

and enter your huggingface token by creating it with write access here

Weight & Biases

The model's loss is logged in the W&B Platform. Provide your API key when prompted; finding it here.


We do not have sudo permissions for writing to /tmp on the supercomputer.

Thus, create the necessary directories, set and export the environment variables, adjust permissions, and add the environment variables to your .bashrc for persistence by doing:

chmod +x setup_wandb.sh
./setup_wandb.sh

Submitting batch jobs

The array mechanism was used for partitioning my tasks into distinct jobs, saving us from having multiple job scripts or continuously adjusting the flags. Look into the job.sub script.


Typically, the job mechanism is utilized to submit multiple similar jobs that run concurrently. This might result in the second job completing faster than the first leading to errors. However, in our setup, each stage relies on the completion of the previous one.


In other words, do not submit more than a single job at once and expect the single submitted job to finish before you execute its successor job.


Submit a single job as:

sbatch --job-name=Stage<task> --array=<task> job.sub

where <task> can be any number from 1 to 11



Suggestion: when you inference, you will see many unicode characters in the output `train.log` file if you access it with a text editor as I am doing `Console(force_terminal=True)` for adding color to the token that replaced the mask token in the output Table that contains the predictions. That said, just do `cat train.log`for seeing the file.

More information on how to run jobs here https://hpcf.cyi.ac.cy/documentation/running_jobs.html

About

Training the first Cypriot Large Language Model on the Masked Language Modeling objective for predicting a given masked word token within a given context

Topics

Resources

Stars

Watchers

Forks