openhermes-mistral-dpo-gptq

This model is a fine-tuned version of TheBloke/OpenHermes-2-Mistral-7B-GPTQ on the None dataset. It achieves the following results on the evaluation set:

• Loss: 1.5203
• Rewards/chosen: 4.8953
• Rewards/rejected: -6.8710
• Rewards/accuracies: 0.875
• Rewards/margins: 11.7663
• Logps/rejected: -512.2221
• Logps/chosen: -446.3599
• Logits/rejected: -2.4853
• Logits/chosen: -2.3998

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: 0.0002
• train_batch_size: 1
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 2
• training_steps: 50
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
0.6384	0.025	10	0.9029	0.3196	-0.5976	0.8125	0.9171	-449.4874	-492.1170	-2.5249	-2.3642
0.0586	0.05	20	1.6495	2.8410	-4.2864	0.8125	7.1274	-486.3763	-466.9029	-2.5061	-2.3959
0.0116	0.075	30	1.5584	4.2892	-6.0162	0.8125	10.3053	-503.6734	-452.4211	-2.4984	-2.4059
1.7733	0.1	40	1.5091	4.7449	-6.6237	0.8125	11.3687	-509.7491	-447.8633	-2.4894	-2.4020
0.0	0.125	50	1.5203	4.8953	-6.8710	0.875	11.7663	-512.2221	-446.3599	-2.4853	-2.3998

Framework versions

• PEFT 0.11.1
• Transformers 4.41.1
• Pytorch 2.0.1+cu117
• Datasets 2.19.1
• Tokenizers 0.19.1