CLUEbenchmark / CLUE

中文语言理解测评基准 Chinese Language Understanding Evaluation Benchmark: datasets, baselines, pre-trained models, corpus and leaderboard
http://www.CLUEbenchmarks.com
4.02k stars 540 forks source link

tnews数据集epoch增大,dev_acc提升,test_acc下降 #136

Open weitajinjucha opened 2 years ago

weitajinjucha commented 2 years ago

tnews数据集epoch增大,dev_acc提升,test_acc下降 --max_seq_length=32 \ --per_gpu_train_batch_size=64 \ --per_gpu_eval_batch_size=64 \ --learning_rate=2e-5 \ --num_train_epochs=5.0 \ --logging_steps=834 \ --save_steps=834 \ epoch设为10时,dev_acc会略微增大,test_acc会显著减小,请问这是什么原因?