dumpmemory/lassl

Easy framework for pre-training language models.

Easy Language Model Pretraining leveraging Huggingface's Transformers and Datasets

What is LASSL • How to Use

English | 한국어

What is LASSL

LASSL is a LAnguage framework for Self-Supervised Learning. LASSL aims to provide an easy-to-use framework for pretraining language model by only using Huggingface's Transformers and Datasets.

Environment setting

First of all, you must install a valid version of pytorch along your computing envrionment. Next, You can install lasslthe required packages to use lassl following.

pip3 install .

How to Use

• Language model pretraining can be divided into three steps: 1. Train Tokenizer, 2. Serialize Corpus, 3.Pretrain Language Model.
• After preparing corpus following to supported corpus type, you can pretrain your own language model.

1. Train Tokenizer

python3 train_tokenizer.py \
    --corpora_dirpath $CORPORA_DIR \
    --corpus_type $CORPUS_TYPE \
    --sampling_ratio $SAMPLING_RATIO \
    --model_type $MODEL_TYPE \
    --vocab_size $VOCAB_SIZE \
    --min_frequency $MIN_FREQUENCY

2. Serialize Corpora

python3 serialize_corpora.py \
    --model_type $MODEL_TYPE \
    --tokenizer_dir $TOKENIZER_DIR \
    --corpora_dir $CORPORA_DIR \
    --corpus_type $CORPUS_TYPE \
    --max_length $MAX_LENGTH \
    --num_proc $NUM_PROC

3. Pretrain Language Model

python3 pretrain_language_model.py --config_path $CONFIG_PATH

# When using TPU, use the command below. (Poetry environment does not provide PyTorch XLA as default.)
python3 xla_spawn.py --num_cores $NUM_CORES pretrain_language_model.py --config_path $CONFIG_PATH

Contributors

Boseop Kim	Minho Ryu	Inje Ryu	Jangwon Park	Hyoungseok Kim
Github	Github	Github	Github	Github

Acknowledgements

LASSL is built with Cloud TPU support from the Tensorflow Research Cloud (TFRC) program.