kant-gpt2

This model is a fine-tuned version of dbmdz/german-gpt2 on an unknown dataset. It achieves the following results on the evaluation set:

• Loss: 1.8022

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: 5e-05
• train_batch_size: 16
• eval_batch_size: 16
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 22

Training results

Training Loss	Epoch	Step	Validation Loss
3.3257	1.0	1825	3.2231
2.9885	2.0	3650	3.0069
2.7955	3.0	5475	2.8440
2.5748	4.0	7300	2.7059
2.3545	5.0	9125	2.5806
2.1759	6.0	10950	2.4618
1.9697	7.0	12775	2.3553
1.7778	8.0	14600	2.2517
1.6192	9.0	16425	2.1599
1.4675	10.0	18250	2.0895
1.3195	11.0	20075	2.0138
1.2012	12.0	21900	1.9602
1.0828	13.0	23725	1.9097
0.9926	14.0	25550	1.8720
0.9076	15.0	27375	1.8426
0.8336	16.0	29200	1.8214
0.7649	17.0	31025	1.8058
0.7208	18.0	32850	1.7980
0.6798	19.0	34675	1.7938
0.647	20.0	36500	1.7969
0.6226	21.0	38325	1.7975
0.601	22.0	40150	1.8022

Framework versions

• Transformers 4.19.2
• Pytorch 1.11.0+cu113
• Tokenizers 0.12.1