Gemma-2B-It-ORPO-SALT

This model is a fine-tuned version of google/gemma-2b-it on the dpo_mix_en and the bct_non_cot_dpo_1000 datasets. It achieves the following results on the evaluation set:

• Loss: 1.3294
• Rewards/chosen: -0.1257
• Rewards/rejected: -0.1400
• Rewards/accuracies: 0.5345
• Rewards/margins: 0.0142
• Logps/rejected: -1.3996
• Logps/chosen: -1.2573
• Logits/rejected: -20.6959
• Logits/chosen: -20.6644
• Sft Loss: 1.2573
• Odds Ratio Loss: 0.7207

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
1.4057	0.8082	500	1.3871	-0.1315	-0.1452	0.5318	0.0137	-1.4522	-1.3148	-21.9402	-21.9144	1.3148	0.7223
1.2196	1.6165	1000	1.3397	-0.1268	-0.1408	0.5355	0.0140	-1.4079	-1.2678	-20.9111	-20.8823	1.2678	0.7187
1.2883	2.4247	1500	1.3294	-0.1257	-0.1400	0.5345	0.0142	-1.3996	-1.2573	-20.6959	-20.6644	1.2573	0.7207

Framework versions

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