opt-babylm2-subset-default-20-epochs-3e-4

This model was trained from scratch on the kanishka/babylm2-subset dataset. It achieves the following results on the evaluation set:

• Loss: 2.4374
• Accuracy: 0.5310

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: 0.0003
• train_batch_size: 32
• eval_batch_size: 64
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 32000
• num_epochs: 20.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy
2.6575	1.0	14169	2.8571	0.4750
2.4179	2.0	28338	2.6258	0.4990
2.2903	3.0	42507	2.5130	0.5119
2.2222	4.0	56676	2.4525	0.5194
2.1477	5.0	70845	2.4207	0.5237
2.0967	6.0	85014	2.3990	0.5266
2.0608	7.0	99183	2.3885	0.5284
2.0231	8.0	113352	2.3815	0.5296
1.9966	9.0	127521	2.3803	0.5304
1.974	10.0	141690	2.3802	0.5311
1.9469	11.0	155859	2.3798	0.5315
1.9241	12.0	170028	2.3857	0.5318
1.8975	13.0	184197	2.3908	0.5319
1.8763	14.0	198366	2.3943	0.5320
1.8519	15.0	212535	2.4029	0.5317
1.8306	16.0	226704	2.4098	0.5317
1.8045	17.0	240873	2.4174	0.5315
1.7801	18.0	255042	2.4237	0.5314
1.7649	19.0	269211	2.4308	0.5311
1.7481	20.0	283380	2.4374	0.5310

Framework versions

• Transformers 4.42.4
• Pytorch 2.2.0+cu121
• Datasets 2.16.1
• Tokenizers 0.19.1