Mistral-7B-Instruct-v0.3-ORPO

This model is a fine-tuned version of mistralai/Mistral-7B-Instruct-v0.3 on the dpo_mix_en dataset. It achieves the following results on the evaluation set:

• Loss: 0.8734
• Rewards/chosen: -0.0810
• Rewards/rejected: -0.1017
• Rewards/accuracies: 0.5720
• Rewards/margins: 0.0208
• Logps/rejected: -1.0175
• Logps/chosen: -0.8098
• Logits/rejected: -3.1455
• Logits/chosen: -3.1171
• Sft Loss: 0.8098
• Odds Ratio Loss: 0.6360

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-06
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• gradient_accumulation_steps: 8
• total_train_batch_size: 16
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 0.1
• num_epochs: 3.0

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen	Sft Loss	Odds Ratio Loss
0.9464	0.8891	500	0.8919	-0.0828	-0.1031	0.5690	0.0202	-1.0306	-0.8281	-3.1432	-3.1149	0.8281	0.6374
0.8737	1.7782	1000	0.8774	-0.0814	-0.1019	0.5760	0.0205	-1.0186	-0.8136	-3.1431	-3.1139	0.8136	0.6371
0.8923	2.6673	1500	0.8734	-0.0810	-0.1017	0.5720	0.0208	-1.0175	-0.8098	-3.1455	-3.1171	0.8098	0.6360

Framework versions

• PEFT 0.10.0
• Transformers 4.40.1
• Pytorch 2.3.0
• Datasets 2.19.0
• Tokenizers 0.19.1