juletx / gpt2-eus

Pretraining GPT2 model on Basque language

Pretraining GPT2 model on Basque

Pretraining code for GPT2 model on Basque language. Includes code to train a tokenizer and a model on various Basque datasets and GPT2 models of different sizes. You can find a base tokenizer and model trained on Euscrawl at https://huggingface.co/HiTZ/gpt2-eus-euscrawl.

Data

There are three choices for pretraining data: Euscrawl, CC100 and mC4. You can add more options.

Tokenizer

The tokenizer is trained on the pretraining data. The tokenizer is trained using the tokenizer/train_tokenizer.py script. The script takes the HuggingFace tokenizer name and dataset names as input. For example, if you want to train the base GPT2 model tokenizer, you can run the following command.

python3 train_tokenizer.py gpt2 HiTZ/euscrawl

The script tokenizer/train_tokenizer.slurm trains a tokenizer for each dataset option. You can modify it to train a tokenizer for other models and datasets.

Model

The model is trained on the pretraining data. The model is trained using the model/run_clm.py script. The script takes a config file as input. The config file contains the model name, tokenizer name, dataset names, and other hyperparameters. For example, if you want to train the base GPT2 model on Euscrawl, you can run the following command.

torchrun --standalone --nproc_per_node=4 run_clm.py ../configs/gpt2-eus-euscrawl.yaml

There are some example configurations in configs directory and some scripts in model directory. For example, the script gpt2-eus-euscrawl.slurm trains a base gpt model on Euscrawl.