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: 0.6796
• Rewards/chosen: 0.0825
• Rewards/rejected: 0.0991
• Rewards/accuracies: 0.375
• Rewards/margins: -0.0166
• Logps/rejected: -111.4488
• Logps/chosen: -104.0037
• Logits/rejected: -1.8100
• Logits/chosen: -1.8966

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: 40
• 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
0.6782	0.01	10	0.6854	0.0505	0.0332	0.5625	0.0173	-112.1082	-104.3235	-1.7988	-1.8929
0.7064	0.01	20	0.6812	0.0509	0.0279	0.8125	0.0230	-112.1610	-104.3192	-1.8032	-1.8968
0.7024	0.01	30	0.6820	0.0697	0.0728	0.375	-0.0031	-111.7118	-104.1311	-1.8068	-1.8953
0.6946	0.02	40	0.6796	0.0825	0.0991	0.375	-0.0166	-111.4488	-104.0037	-1.8100	-1.8966

Framework versions

• Transformers 4.35.2
• Pytorch 2.0.1+cu117
• Datasets 2.16.1
• Tokenizers 0.15.0