Gemma-2B-It-ORPO

This model is a fine-tuned version of google/gemma-2b-it on the dpo_mix_en dataset. It achieves the following results on the evaluation set:

• Loss: 1.3996
• Rewards/chosen: -0.1332
• Rewards/rejected: -0.1501
• Rewards/accuracies: 0.5460
• Rewards/margins: 0.0169
• Logps/rejected: -1.5010
• Logps/chosen: -1.3319
• Logits/rejected: -20.7443
• Logits/chosen: -20.6555
• Sft Loss: 1.3319
• Odds Ratio Loss: 0.6769

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
1.5296	0.8891	500	1.4462	-0.1378	-0.1548	0.5470	0.0170	-1.5481	-1.3781	-21.5745	-21.4875	1.3781	0.6806
1.4761	1.7782	1000	1.4072	-0.1340	-0.1510	0.5350	0.0170	-1.5095	-1.3395	-20.8914	-20.8038	1.3395	0.6771
1.4105	2.6673	1500	1.3996	-0.1332	-0.1501	0.5460	0.0169	-1.5010	-1.3319	-20.7443	-20.6555	1.3319	0.6769

Framework versions

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