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: 0.9029
• Rewards/chosen: -0.1592
• Rewards/rejected: -0.0751
• Rewards/accuracies: 0.4375
• Rewards/margins: -0.0841
• Logps/rejected: -164.9728
• Logps/chosen: -207.7616
• Logits/rejected: -2.4937
• Logits/chosen: -2.5880

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
0.6777	0.005	10	0.7136	-0.0128	-0.0514	0.5625	0.0386	-164.7356	-206.2971	-2.4991	-2.5809
0.6983	0.01	20	0.7209	-0.0223	-0.1062	0.625	0.0838	-165.2831	-206.3929	-2.4962	-2.5809
0.697	0.015	30	0.7341	-0.0064	-0.0583	0.6875	0.0519	-164.8043	-206.2330	-2.4984	-2.5864
0.6967	0.02	40	0.7473	-0.0052	-0.0485	0.4375	0.0433	-164.7064	-206.2214	-2.4930	-2.5857
0.6666	0.025	50	0.9029	-0.1592	-0.0751	0.4375	-0.0841	-164.9728	-207.7616	-2.4937	-2.5880

Framework versions

• PEFT 0.12.0
• Transformers 4.42.4
• Pytorch 2.0.1+cu117
• Datasets 2.20.0
• Tokenizers 0.19.1