How come the IJB B,C results are better than published in paper?

Hi, thanks a lot for your outstanding work.

I downloaded your shared data sets, pretrained model, and ran the evaluation and got the following results,

python -u IJB_11.py --model-prefix ./pretrained_models/r100-arcface/model --model-epoch 0 --gpu 0 --target IJBB --job arcface
+--------------+-------+-------+--------+-------+-------+-------+
| Methods | 1e-06 | 1e-05 | 0.0001 | 0.001 | 0.01 | 0.1 |
+--------------+-------+-------+--------+-------+-------+-------+
| arcface-IJBB | 42.75 | 90.76 | 94.75 | 96.48 | 97.64 | 98.67 |
+--------------+-------+-------+--------+-------+-------+-------+

python -u IJB_11.py --model-prefix ./pretrained_models/r100-arcface/model --model-epoch 0 --gpu 0 --target IJBC --job arcface
+--------------+-------+-------+--------+-------+-------+-------+
| Methods | 1e-06 | 1e-05 | 0.0001 | 0.001 | 0.01 | 0.1 |
+--------------+-------+-------+--------+-------+-------+-------+
| arcface-IJBC | 89.91 | 94.47 | 96.28 | 97.52 | 98.36 | 99.08 |
+--------------+-------+-------+--------+-------+-------+-------+

At FAR=1E-4, the TARs are 0.9475, and 0.9628 .
In your ArcFace paper https://arxiv.org/pdf/1801.07698.pdf, in Table 7, you had at FAR=1E-4,
MS1MV2, R100, ArcFace 0.942 0.956

The results I got with your evaluation code are 0.05 better than your paper. So what has changed since you published your paper?