search

ylsung / Ladder-Side-Tuning

PyTorch codes for "LST: Ladder Side-Tuning for Parameter and Memory Efficient Transfer Learning"
MIT License
230 stars 10 forks source link

Scores of individual GLUE tasks for BART encoder #11

Closed BaohaoLiao closed 1 year ago

BaohaoLiao commented 1 year ago

Hi,

in Table 2 you used BART-large encoder. Is it possible for you to release the score of each individual GLUE task instead of the average one?

ylsung commented 1 year ago

Hi Baolao,


Thanks for your interest.

The per-task performance in Table 2 for our approach is 

cola sst2 mrpc qqp mnli-m mnli-mm qnli rte avg
55.5 94.72 87.25 89.47 85.97 86.31 91.93 69.68 82.09857143


We follow the Y-tuning paper to report the best result out of different learning rates (3e-4, 1e-3, 3e-3) and three seeds. The best learning rate for QQP and MNLI is 3e-4 and for the rest of the tasks is 1e-3.

Best,
Yi-Lin  

BaohaoLiao commented 1 year ago

Thank you for your results!!!