mistral-7b-dpo-full-sft-wo-medication_qa

This model is a fine-tuned version of Minbyul/mistral-7b-wo-medication_qa-sft on the HuggingFaceH4/ultrafeedback_binarized dataset. It achieves the following results on the evaluation set:

• Loss: 0.0756
• Rewards/chosen: -3.7115
• Rewards/rejected: -11.1989
• Rewards/accuracies: 0.9531
• Rewards/margins: 7.4875
• Logps/rejected: -1662.8185
• Logps/chosen: -803.2770
• Logits/rejected: -2.3910
• Logits/chosen: -2.5860

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: 8
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 2
• total_train_batch_size: 64
• total_eval_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 1

Training results

Training Loss	Epoch	Step	Logits/chosen	Logits/rejected	Logps/chosen	Logps/rejected	Validation Loss	Rewards/accuracies	Rewards/chosen	Rewards/margins	Rewards/rejected
0.2799	0.31	100	-3.0348	-3.0868	-584.1479	-794.0103	0.5261	0.75	-1.5202	0.9907	-2.5108
0.154	0.62	200	-2.6948	-2.5547	-742.1359	-1446.8754	0.0923	0.9375	-3.1001	5.9394	-9.0395
0.0948	0.92	300	-2.5877	-2.3930	-803.1033	-1661.4266	0.0753	0.9531	-3.7097	7.4753	-11.1850

Framework versions

• Transformers 4.39.0.dev0
• Pytorch 2.1.2
• Datasets 2.14.6
• Tokenizers 0.15.2