opt-babylm2-rewritten-clean-spacy-32k-earlystop-40epochs_seed-42_3e-4

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

• Loss: 2.9892
• Accuracy: 0.4202

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
• gradient_accumulation_steps: 8
• total_train_batch_size: 256
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 32000
• num_epochs: 40.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy
6.8783	0.9996	1931	4.4654	0.2891
4.2263	1.9997	3863	3.9179	0.3337
3.7724	2.9999	5795	3.6397	0.3562
3.5091	4.0	7727	3.4569	0.3724
3.3306	4.9996	9658	3.3310	0.3838
3.2012	5.9997	11590	3.2469	0.3918
3.1088	6.9999	13522	3.1828	0.3982
3.0364	8.0	15454	3.1404	0.4023
2.9837	8.9996	17385	3.1080	0.4057
2.9377	9.9997	19317	3.0840	0.4077
2.9019	10.9999	21249	3.0633	0.4101
2.8713	12.0	23181	3.0505	0.4117
2.8449	12.9996	25112	3.0376	0.4130
2.8231	13.9997	27044	3.0270	0.4143
2.7828	14.9999	28976	3.0222	0.4150
2.7644	16.0	30908	3.0160	0.4156
2.7508	16.9996	32839	3.0100	0.4167
2.7296	17.9997	34771	3.0023	0.4178
2.7053	18.9999	36703	2.9967	0.4188
2.6821	20.0	38635	2.9926	0.4195
2.6601	20.9996	40566	2.9892	0.4202
2.6406	21.9997	42498	2.9898	0.4204
2.621	22.9999	44430	2.9921	0.4205
2.6048	24.0	46362	2.9928	0.4207

Framework versions

• Transformers 4.45.1
• Pytorch 2.4.1+cu121
• Datasets 3.0.1
• Tokenizers 0.20.0