mega-small-2048-C1024-tk_id-simplewiki-MR50

mega-small-2048 on simple wikipedia

MEGA for masked LM 'small' (12 layers, 512 hidden size, 2048 ctx in chunks of 1024) on the pszemraj/simple_wikipedia_LM dataset. It achieves the following results on the evaluation set:

• Loss: 3.4773
• Accuracy: 0.4591

Model description

See config for architecture details. While not a ready 'pretrained' model, this was trained from scratch.

This model uses the tokenizer from roberta-base.

Intended uses & limitations

More information needed

Training and evaluation data

Note: this was trained in bf16. the official recommendation is fp32 - still exploring this.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0005
• train_batch_size: 1
• eval_batch_size: 1
• seed: 3208
• gradient_accumulation_steps: 64
• total_train_batch_size: 64
• 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

Additionally:

• mask rate of 50% (See paper for details)
• whole-word masking

Training results

Framework versions

• Transformers 4.33.1
• Pytorch 2.2.0.dev20230907+cu118
• Datasets 2.13.1
• Tokenizers 0.13.3