smolm-autoreg-bpe-counterfactual_babylm_measure_nps_as_singular_new-3e-4

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

• Loss: 3.4141
• Accuracy: 0.4094

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
3.7385	1.0	18602	3.9139	0.3449
3.4266	2.0	37204	3.6133	0.3747
3.2882	3.0	55806	3.4788	0.3899
3.2035	4.0	74408	3.4080	0.3974
3.141	5.0	93010	3.4152	0.3997
3.0971	6.0	111612	3.3875	0.4027
3.0545	7.0	130214	3.3727	0.4048
3.0227	8.0	148816	3.3644	0.4070
2.9947	9.0	167418	3.3735	0.4077
2.966	10.0	186020	3.3659	0.4086
2.9448	11.0	204622	3.3568	0.4092
2.9212	12.0	223224	3.3682	0.4092
2.8974	13.0	241826	3.3732	0.4089
2.8755	14.0	260428	3.3812	0.4096
2.8571	15.0	279030	3.3808	0.4100
2.8333	16.0	297632	3.3887	0.4097
2.8156	17.0	316234	3.3931	0.4098
2.7983	18.0	334836	3.3972	0.4099
2.778	19.0	353438	3.4080	0.4095
2.7615	20.0	372040	3.4141	0.4094

Framework versions

• Transformers 4.38.0
• Pytorch 2.3.1+cu121
• Datasets 2.16.1
• Tokenizers 0.15.2