electra_large_discriminator_squad2_512

python ${EXAMPLES}/run_squad.py \
--model_type electra \
--model_name_or_path google/electra-large-discriminator \
--do_train \
--do_eval \
--train_file ${SQUAD}/train-v2.0.json \
--predict_file ${SQUAD}/dev-v2.0.json \
--version_2_with_negative \
--do_lower_case \
--num_train_epochs 3 \
--warmup_steps 306 \
--weight_decay 0.01 \
--learning_rate 3e-5 \
--max_grad_norm 0.5 \
--adam_epsilon 1e-6 \
--max_seq_length 512 \
--doc_stride 128 \
--per_gpu_train_batch_size 8 \
--gradient_accumulation_steps 16 \
--per_gpu_eval_batch_size 128 \
--fp16 \
--fp16_opt_level O1 \
--threads 12 \
--logging_steps 50 \
--save_steps 1000 \
--overwrite_output_dir \
--output_dir ${MODEL_PATH}

El modelo utiliza las siguientes versiones de software y plataforma:

• Transformers: 2.11.0
• PyTorch: 1.5.0
• TensorFlow: 2.2.0
• Python: 3.8.1
• OS/Plataforma: Linux-5.3.0-59-generic-x86_64-with-glibc2.10
• CPU/GPU: Intel i9-9900K / NVIDIA Titan RTX 24GB