conll03.sh can't reproduce the f1 score in paper

[Senwang98]

@xiaoya-li
Hi, I used conll03.sh and bert-large-cased to reproduce your f1 score in acl20 paper, but failed.
here is my config:

#!/usr/bin/env bash
# -*- coding: utf-8 -*-


TIME=0901
FILE=conll03_cased_large
REPO_PATH=/opt/tiger/ws/mrc-for-flat-nested-ner
export PYTHONPATH="$PYTHONPATH:$REPO_PATH"

DATA_DIR=/opt/tiger/ws/ner_dataset/conll03
BERT_DIR=/opt/tiger/ws/bert-large-cased
OUTPUT_BASE=${REPO_PATH}/outputs

BATCH=10
GRAD_ACC=4
BERT_DROPOUT=0.1
MRC_DROPOUT=0.3
LR=3e-5
LR_MINI=3e-7
LR_SCHEDULER=polydecay
SPAN_WEIGHT=0.1
WARMUP=0
MAX_LEN=200
MAX_NORM=1.0
MAX_EPOCH=20
INTER_HIDDEN=2048
WEIGHT_DECAY=0.01
OPTIM=torch.adam
VAL_CHECK=0.2
PREC=16
SPAN_CAND=pred_and_gold


OUTPUT_DIR=${OUTPUT_BASE}/mrc_ner/${FILE}_cased_large_lr${LR}_drop${MRC_DROPOUT}_norm${MAXNORM}_weight${SPAN_WEIGHT}_warmup${WARMUP}_maxlen${MAXLEN}
mkdir -p ${OUTPUT_DIR}


CUDA_VISIBLE_DEVICES=0,1 python ${REPO_PATH}/train/mrc_ner_trainer.py \
--data_dir ${DATA_DIR} \
--bert_config_dir ${BERT_DIR} \
--max_length ${MAX_LEN} \
--batch_size ${BATCH} \
--gpus="2" \
--precision=${PREC} \
--progress_bar_refresh_rate 1 \
--lr ${LR} \
--val_check_interval ${VAL_CHECK} \
--accumulate_grad_batches ${GRAD_ACC} \
--default_root_dir ${OUTPUT_DIR} \
--mrc_dropout ${MRC_DROPOUT} \
--bert_dropout ${BERT_DROPOUT} \
--max_epochs ${MAX_EPOCH} \
--span_loss_candidates ${SPAN_CAND} \
--weight_span ${SPAN_WEIGHT} \
--warmup_steps ${WARMUP} \
--distributed_backend=ddp \
--max_length ${MAX_LEN} \
--gradient_clip_val ${MAX_NORM} \
--weight_decay ${WEIGHT_DECAY} \
--optimizer ${OPTIM} \
--lr_scheduler ${LR_SCHEDULER} \
--classifier_intermediate_hidden_size ${INTER_HIDDEN} \
--flat \
--lr_mini ${LR_MINI}

and the final test f1 result:

Testing: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████▊| 1380/1382 [01:06<00:00, 21.21it/s]

TEST INFO -> test_f1 is: 0.8897400498390198 precision: 0.8865866661071777, recall: 0.8929159641265869

do you have any suggestionns?