mistral-dpo

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: 0.0001
• Rewards/chosen: 2.5279
• Rewards/rejected: -6.8729
• Rewards/accuracies: 1.0
• Rewards/margins: 9.4009
• Logps/rejected: -86.5415
• Logps/chosen: -10.6380
• Logits/rejected: -2.2909
• Logits/chosen: -2.3250

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
0.5916	0.09	10	0.3053	0.6309	-0.4316	1.0	1.0625	-22.1280	-29.6084	-2.4417	-2.4765
0.127	0.19	20	0.0029	1.9393	-4.1219	1.0	6.0612	-59.0316	-16.5245	-2.3738	-2.4158
0.0013	0.28	30	0.0003	2.3840	-5.7860	1.0	8.1700	-75.6720	-12.0770	-2.3067	-2.3466
0.0002	0.37	40	0.0001	2.4704	-6.5625	1.0	9.0328	-83.4367	-11.2135	-2.2895	-2.3248
0.0002	0.46	50	0.0001	2.5279	-6.8729	1.0	9.4009	-86.5415	-10.6380	-2.2909	-2.3250

Framework versions

• Transformers 4.35.2
• Pytorch 2.0.1+cu117
• Datasets 2.15.0
• Tokenizers 0.15.0