Gemma-7B-It-ORPO-SALT

This model is a fine-tuned version of google/gemma-7b-it on the dpo_mix_en and the bct_non_cot_dpo_1000 datasets. It achieves the following results on the evaluation set:

• Loss: 1.2657
• Rewards/chosen: -0.1198
• Rewards/rejected: -0.1438
• Rewards/accuracies: 0.5700
• Rewards/margins: 0.0239
• Logps/rejected: -1.4377
• Logps/chosen: -1.1983
• Logits/rejected: 253.9599
• Logits/chosen: 253.6037
• Sft Loss: 1.1983
• Odds Ratio Loss: 0.6746

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
1.374	0.8082	500	1.3436	-0.1276	-0.1503	0.5673	0.0227	-1.5033	-1.2762	249.9064	249.6123	1.2762	0.6738
1.1628	1.6165	1000	1.2833	-0.1215	-0.1446	0.5618	0.0231	-1.4461	-1.2153	253.1810	252.8272	1.2153	0.6796
1.1874	2.4247	1500	1.2657	-0.1198	-0.1438	0.5700	0.0239	-1.4377	-1.1983	253.9599	253.6037	1.1983	0.6746

Framework versions

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