taskydata/deberta-v3-base_10xp3nirstbbflanseuni_10xc4

Hyperparameters:

• learning rate: 2e-5
• weight decay: 0.01
• per_device_train_batch_size: 8
• per_device_eval_batch_size: 8
• gradient_accumulation_steps:1
• eval steps: 50000
• max_length: 512
• num_epochs: 1
• hidden_dropout_prob: 0.3
• attention_probs_dropout_prob: 0.25

Dataset version:

• tasky_or_not/10xp3nirstbbflanseuni_10xc4

Checkpoint:

• 300000 steps.

Results on Validation set:

Step	Training Loss	Validation Loss	Accuracy	Precision	Recall	F1
50000	0.020800	0.192550	0.970363	0.990686	0.949654	0.969736
100000	0.015200	0.264168	0.969427	0.994374	0.944196	0.968636
150000	0.012900	0.146541	0.981440	0.994599	0.968138	0.981190
200000	0.011100	0.319310	0.970516	0.998871	0.942097	0.969654
250000	0.008000	0.204103	0.976309	0.996226	0.956241	0.975824
300000	0.006100	0.096262	0.988053	0.994676	0.981358	0.987972
350000	0.005800	0.162989	0.983663	0.994730	0.972478	0.983478

Wandb logs:

• https://wandb.ai/manandey/taskydata/runs/y3j3fbkh?workspace=user-manandey