smolm-autoreg-bpe-counterfactual-babylm-adj_num_freq_balanced-3e-4

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

• Loss: 3.4663
• Accuracy: 0.4052

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
3.7291	1.0	18629	3.9229	0.3464
3.4237	2.0	37258	3.6542	0.3747
3.2842	3.0	55887	3.5183	0.3880
3.1952	4.0	74516	3.4748	0.3938
3.1351	5.0	93145	3.4493	0.3972
3.0844	6.0	111774	3.4156	0.4005
3.0442	7.0	130403	3.3854	0.4032
3.008	8.0	149032	3.4062	0.4030
2.9768	9.0	167661	3.3970	0.4047
2.9498	10.0	186290	3.4024	0.4047
2.917	11.0	204919	3.4242	0.4039
2.9005	12.0	223548	3.4093	0.4049
2.8747	13.0	242177	3.4192	0.4051
2.8542	14.0	260806	3.4233	0.4053
2.8326	15.0	279435	3.4314	0.4054
2.8125	16.0	298064	3.4404	0.4052
2.7911	17.0	316693	3.4450	0.4054
2.7682	18.0	335322	3.4488	0.4054
2.7512	19.0	353951	3.4581	0.4054
2.7331	20.0	372580	3.4663	0.4052

Framework versions

• Transformers 4.36.0
• Pytorch 2.1.1+cu121
• Datasets 2.15.0
• Tokenizers 0.15.0