Can not reproduce the result of Roberta-Base about lora

[jcluo-2022]

I also met this question,

Specifcally, I downloaded the LoRa weights and run the following commands.

python -m torch.distributed.launch-
-nproc_per_node=1
examples/text-classification/run_glue.py
--model_name_or_path roberta-base
--lora_path ./cp/roberta_base_lora_mnli.bin
--task_name mnli
--do_eval
--output_dir ./output
--apply_lora
--lora_r 8
--lora_alpha 8

It turns out that the accuracy is 0.6655 while the number in paper is 0.875

[Car-pe]

Hello, I guess you used the wrong hyper parameters since the gap is really large.
But I also can not reproduce the result, the paper claimed the STSB has achieved 91.5, but my result is 91.03.

And for Roberta-large, I also can not reproduce the results.
Cola: 68.2(paper) - 67.0(my), gap is 1.2
Stab: 92.6(paper) - 92.2(my), gap is 0.4
Mrpc:90.9(paper) - 90.7(my), gap is 0.2