zephyr0-7b-ultra-dpo

This model is a fine-tuned version of HuggingFaceH4/mistral-7b-sft-beta on an unknown dataset. It achieves the following results on the evaluation set:

• Loss: 0.5302
• Rewards/chosen: -1.0376
• Rewards/rejected: -1.8238
• Rewards/accuracies: 0.7422
• Rewards/margins: 0.7863
• Logps/rejected: -410.1681
• Logps/chosen: -326.1224
• Logits/rejected: -1.9724
• Logits/chosen: -2.0723

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: 5e-07
• train_batch_size: 2
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 8
• total_train_batch_size: 128
• total_eval_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 1.0

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
0.5846	0.2060	100	0.5656	-0.7501	-1.1978	0.6797	0.4477	-347.5626	-297.3736	-2.4976	-2.5597
0.5397	0.4119	200	0.5447	-1.2075	-1.8283	0.7344	0.6208	-410.6155	-343.1137	-2.1704	-2.2467
0.5186	0.6179	300	0.5416	-0.9175	-1.6245	0.75	0.7071	-390.2354	-314.1128	-2.0885	-2.1736
0.5143	0.8239	400	0.5321	-1.0168	-1.7802	0.7344	0.7634	-405.8048	-324.0463	-2.0197	-2.1154

Framework versions

• Transformers 4.45.1
• Pytorch 2.4.1+cu121
• Datasets 3.0.0
• Tokenizers 0.20.0