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: 3.2500
• Rewards/chosen: -1.0975
• Rewards/rejected: -1.6306
• Rewards/accuracies: 0.625
• Rewards/margins: 0.5331
• Logps/rejected: -307.3866
• Logps/chosen: -331.8629
• Logits/rejected: -2.4077
• Logits/chosen: -2.3038

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: 300
• 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
4.2921	0.03	50	9.8028	-5.3862	0.1060	0.1875	-5.4922	-290.0201	-374.7499	-2.2861	-2.1795
9.75	0.05	100	8.8191	-12.7493	-8.6505	0.3125	-4.0989	-377.5849	-448.3811	-2.2836	-2.2309
3.2104	0.07	150	0.8915	-3.5710	-6.0350	0.375	2.4640	-351.4305	-356.5982	-2.6543	-2.5955
2.655	0.1	200	0.3207	-1.0209	-4.6027	0.6875	3.5818	-337.1074	-331.0971	-2.4341	-2.3534
4.8481	0.12	250	1.1311	-0.8147	-2.3072	0.625	1.4926	-314.1525	-329.0346	-2.3257	-2.2374
3.1598	0.15	300	3.2500	-1.0975	-1.6306	0.625	0.5331	-307.3866	-331.8629	-2.4077	-2.3038

Framework versions

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