smolm-autoreg-bpe-counterfactual_babylm_naans_new-seed_1024-1e-4

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

• Loss: 3.4373
• Accuracy: 0.4072

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: 0.0001
• train_batch_size: 32
• eval_batch_size: 64
• seed: 1024
• 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
4.0553	1.0	18595	4.2752	0.3077
3.5587	2.0	37190	3.7501	0.3633
3.3865	3.0	55785	3.5884	0.3815
3.2906	4.0	74380	3.5054	0.3891
3.2202	5.0	92975	3.4769	0.3943
3.1699	6.0	111570	3.4426	0.3977
3.1232	7.0	130165	3.4478	0.3994
3.091	8.0	148760	3.4243	0.4014
3.0613	9.0	167355	3.4169	0.4030
3.0322	10.0	185950	3.4142	0.4050
3.0107	11.0	204545	3.4052	0.4049
2.9844	12.0	223140	3.4128	0.4053
2.9671	13.0	241735	3.4150	0.4062
2.9477	14.0	260330	3.4174	0.4062
2.9272	15.0	278925	3.4275	0.4067
2.9088	16.0	297520	3.4271	0.4068
2.8915	17.0	316115	3.4245	0.4071
2.872	18.0	334710	3.4262	0.4070
2.8514	19.0	353305	3.4322	0.4073
2.8467	20.0	371900	3.4373	0.4072

Framework versions

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