Gemma-7B-It-ORPO-SALT-HALF

This model is a fine-tuned version of google/gemma-7b-it on the dpo_mix_en and the bct_non_cot_dpo_500 datasets. It achieves the following results on the evaluation set:

• Loss: 1.3159
• Rewards/chosen: -0.1249
• Rewards/rejected: -0.1471
• Rewards/accuracies: 0.5619
• Rewards/margins: 0.0222
• Logps/rejected: -1.4709
• Logps/chosen: -1.2488
• Logits/rejected: 253.8645
• Logits/chosen: 253.5439
• Sft Loss: 1.2488
• Odds Ratio Loss: 0.6713

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
1.422	0.8467	500	1.3896	-0.1322	-0.1546	0.5752	0.0224	-1.5459	-1.3222	250.5634	250.2739	1.3222	0.6733
1.3103	1.6935	1000	1.3313	-0.1264	-0.1489	0.5695	0.0224	-1.4886	-1.2642	253.1350	252.8147	1.2642	0.6718
1.2057	2.5402	1500	1.3159	-0.1249	-0.1471	0.5619	0.0222	-1.4709	-1.2488	253.8645	253.5439	1.2488	0.6713

Framework versions

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