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: 10.4812
• Rewards/chosen: -34.3483
• Rewards/rejected: -10.5927
• Rewards/accuracies: 0.4375
• Rewards/margins: -23.7556
• Logps/rejected: -369.0945
• Logps/chosen: -652.6519
• Logits/rejected: -1.9434
• Logits/chosen: -2.1850

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

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
0.785	0.01	10	0.6767	0.0342	0.0037	0.5625	0.0305	-263.1300	-308.8264	-2.4042	-2.7773
0.988	0.01	20	0.6634	0.1443	-0.0047	0.75	0.1490	-263.2147	-307.7260	-2.4189	-2.7870
0.9481	0.01	30	10.5125	-35.0203	-10.9252	0.4375	-24.0951	-372.4191	-659.3712	-1.9468	-2.1781
10.47	0.02	40	17.8422	-55.0969	-26.2159	0.375	-28.8810	-525.3258	-860.1373	-1.6618	-1.8474
2.5456	0.03	50	10.4812	-34.3483	-10.5927	0.4375	-23.7556	-369.0945	-652.6519	-1.9434	-2.1850

Framework versions

• Transformers 4.34.1
• Pytorch 2.0.1+cu117
• Datasets 2.14.6
• Tokenizers 0.14.1