XLM-R Longformer Model
XLM-R Longformer is a XLM-R model, that has been extended to allow sequence lengths up to 4096 tokens, instead of the regular 512. The model was pre-trained from the XLM-RoBERTa checkpoint using the Longformer pre-training scheme on the English WikiText-103 corpus.
The reason for this was to investigate methods for creating efficient Transformers for low-resource languages, such as Swedish, without the need to pre-train them on long-context datasets in each respecitve language. The trained model came as a result of a master thesis project at Peltarion and was fine-tuned on multilingual quesion-answering tasks, with code available here.
Since both XLM-R model and Longformer models are large models, it it recommended to run the models with NVIDIA Apex (16bit precision), large GPU and several gradient accumulation steps.
How to Use
The model can be used as expected to fine-tune on a downstream task.
For instance for QA.
import torch
from transformers import AutoModel, AutoTokenizer
MAX_SEQUENCE_LENGTH = 4096
MODEL_NAME_OR_PATH = "markussagen/xlm-roberta-longformer-base-4096"
tokenizer = AutoTokenizer.from_pretrained(
MODEL_NAME_OR_PATH,
max_length=MAX_SEQUENCE_LENGTH,
padding="max_length",
truncation=True,
)
model = AutoModelForQuestionAnswering.from_pretrained(
MODEL_NAME_OR_PATH,
max_length=MAX_SEQUENCE_LENGTH,
)
Training Procedure
The model have been trained on the WikiText-103 corpus, using a 48GB GPU with the following training script and parameters. The model was pre-trained for 6000 iterations and took ~5 days. See the full training script and Github repo for more information
wget https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-103-raw-v1.zip
unzip wikitext-103-raw-v1.zip
export DATA_DIR=./wikitext-103-raw
scripts/run_long_lm.py \
--model_name_or_path xlm-roberta-base \
--model_name xlm-roberta-to-longformer \
--output_dir ./output \
--logging_dir ./logs \
--val_file_path $DATA_DIR/wiki.valid.raw \
--train_file_path $DATA_DIR/wiki.train.raw \
--seed 42 \
--max_pos 4096 \
--adam_epsilon 1e-8 \
--warmup_steps 500 \
--learning_rate 3e-5 \
--weight_decay 0.01 \
--max_steps 6000 \
--evaluate_during_training \
--logging_steps 50 \
--eval_steps 50 \
--save_steps 6000 \
--max_grad_norm 1.0 \
--per_device_eval_batch_size 2 \
--per_device_train_batch_size 1 \
--gradient_accumulation_steps 64 \
--overwrite_output_dir \
--fp16 \
--do_train \
--do_eval