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

This model is a fine-tuned version of mistralai/Mistral-7B-Instruct-v0.3 on the dpo_mix_en and the bct_non_cot_dpo_500 datasets. It achieves the following results on the evaluation set:

• Loss: 0.8506
• Rewards/chosen: -0.0787
• Rewards/rejected: -0.0996
• Rewards/accuracies: 0.5724
• Rewards/margins: 0.0209
• Logps/rejected: -0.9956
• Logps/chosen: -0.7867
• Logits/rejected: -3.1507
• Logits/chosen: -3.1305
• Sft Loss: 0.7867
• Odds Ratio Loss: 0.6382

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
0.8758	0.8467	500	0.8691	-0.0805	-0.1009	0.5705	0.0203	-1.0086	-0.8054	-3.1276	-3.1089	0.8054	0.6371
0.8098	1.6935	1000	0.8549	-0.0791	-0.0999	0.5676	0.0207	-0.9985	-0.7911	-3.1170	-3.0966	0.7911	0.6375
0.8135	2.5402	1500	0.8506	-0.0787	-0.0996	0.5724	0.0209	-0.9956	-0.7867	-3.1507	-3.1305	0.7867	0.6382

Framework versions

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