Vicuna-7B-v1.5-ORPO-SALT

This model is a fine-tuned version of lmsys/vicuna-7b-v1.5 on the dpo_mix_en and the bct_non_cot_dpo_1000 datasets. It achieves the following results on the evaluation set:

• Loss: 0.9497
• Rewards/chosen: -0.0879
• Rewards/rejected: -0.0995
• Rewards/accuracies: 0.5164
• Rewards/margins: 0.0116
• Logps/rejected: -0.9948
• Logps/chosen: -0.8787
• Logits/rejected: -0.3581
• Logits/chosen: -0.3775
• Sft Loss: 0.8787
• Odds Ratio Loss: 0.7104

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
1.0008	0.8082	500	0.9777	-0.0907	-0.1019	0.5055	0.0113	-1.0193	-0.9066	-0.3689	-0.3878	0.9066	0.7105
0.8458	1.6165	1000	0.9560	-0.0885	-0.1000	0.5191	0.0115	-1.0000	-0.8850	-0.3578	-0.3772	0.8850	0.7097
0.9219	2.4247	1500	0.9497	-0.0879	-0.0995	0.5164	0.0116	-0.9948	-0.8787	-0.3581	-0.3775	0.8787	0.7104

Framework versions

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