can not reproduce the the result of roberta large on dataste sst-2

[itongggg]

my transformers version == 4.28.1
i run the the following command
TASK=SST-2 K=16 SEED=42 BS=64 LR=1e-6 EPS=1e-3 MODEL=roberta-large bash mezo.sh
get the test acc 70.2 which in your paper is 89.6 (1.2)
i did not modify any scripts or files, should i ?

[gaotianyu1350]

Can you try the full grid lr = {1e−7, 1e−6, 1e−5} (Table 15 from our paper)?

[shawnnn3di]

I can't reproduce the result either. However after degrading my Python, PyTorch and Transformers packages the results are good. Weird.