mega-ar-large-2048-simplewiki

This is a 'large' size autoregressive MEGA model initialized from random weights and trained on pszemraj/simple_wikipedia_LM for three epochs.

It achieves the following results on the evaluation set:

• Loss: 3.3412
• Accuracy: 0.4360

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0005
• train_batch_size: 1
• eval_batch_size: 1
• seed: 80085
• gradient_accumulation_steps: 32
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.98) and epsilon=1e-07
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.05
• num_epochs: 3.0

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy
7.2245	0.11	100	6.9372	0.0711
6.6575	0.22	200	6.2335	0.1853
5.9406	0.34	300	5.3724	0.2635
5.4452	0.45	400	4.9243	0.2940
5.2524	0.56	500	4.6568	0.3172
4.7862	0.67	600	4.4488	0.3347
4.7132	0.79	700	4.2699	0.3481
4.6601	0.9	800	4.1502	0.3582
4.5067	1.01	900	4.0461	0.3681
4.4465	1.12	1000	3.9488	0.3773
4.4493	1.24	1100	3.8681	0.3833
4.3136	1.35	1200	3.8039	0.3897
4.2978	1.46	1300	3.7373	0.3956
4.0475	1.57	1400	3.6874	0.4003
4.1328	1.68	1500	3.6339	0.4061
4.0758	1.8	1600	3.5866	0.4115
3.8489	1.91	1700	3.5438	0.4163
3.913	2.02	1800	3.5136	0.4192
3.7746	2.13	1900	3.4860	0.4226
3.9547	2.25	2000	3.4505	0.4255
3.9726	2.36	2100	3.4283	0.4269
3.7546	2.47	2200	3.3999	0.4298
3.7442	2.58	2300	3.3820	0.4317
3.6848	2.7	2400	3.3687	0.4333
3.5491	2.81	2500	3.3531	0.4349
3.9563	2.92	2600	3.3412	0.4360

Framework versions

• Transformers 4.33.1
• Pytorch 2.2.0.dev20230907+cu118
• Datasets 2.14.5
• Tokenizers 0.13.3