6.7b-ri-reproduce-combined-4-gpu-20-val

This model is a fine-tuned version of facebook/opt-6.7b on the None dataset. It achieves the following results on the evaluation set:

• Loss: 3.9434
• Accuracy: 0.0329
• Perplexity: 51.5916

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: 9e-07
• train_batch_size: 1
• eval_batch_size: 8
• seed: 100
• distributed_type: multi-GPU
• num_devices: 4
• total_train_batch_size: 4
• total_eval_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: constant
• num_epochs: 15.0

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy	Perplexity
2.5731	1.0	79	2.6113	0.0317	13.6171
2.206	2.0	158	2.4805	0.0328	11.9469
1.9105	3.0	237	2.4512	0.0333	11.6019
1.6301	4.0	316	2.5078	0.0345	12.2780
1.3733	5.0	395	2.6816	0.0342	14.6090
1.1337	6.0	474	3.0078	0.0330	20.2431
0.9619	7.0	553	3.1777	0.0330	23.9923
0.798	8.0	632	3.2559	0.0330	25.9419
0.6653	9.0	711	3.4277	0.0331	30.8068
0.552	10.0	790	3.5566	0.0333	35.0453
0.4568	11.0	869	3.7324	0.0324	41.7802
0.3756	12.0	948	3.8184	0.0328	45.5295
0.3119	13.0	1027	3.8477	0.0331	46.8831
0.2448	14.0	1106	3.9062	0.0329	49.7122
0.1986	15.0	1185	3.9434	0.0329	51.5916

Framework versions

• Transformers 4.25.0.dev0
• Pytorch 1.12.1+cu113
• Datasets 2.3.2
• Tokenizers 0.12.1