gemma-7b-orpo

This model is a fine-tuned version of google/gemma-7b on the argilla/dpo-mix-7k dataset. It achieves the following results on the evaluation set:

• Loss: 1.7559
• Rewards/chosen: -0.0650
• Rewards/rejected: -0.0764
• Rewards/accuracies: 0.5971
• Rewards/margins: 0.0114
• Logps/rejected: -1.5282
• Logps/chosen: -1.3004
• Logits/rejected: 266.0260
• Logits/chosen: 295.6202
• Nll Loss: 1.6941
• Log Odds Ratio: -0.6992
• Log Odds Chosen: 0.3721

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: 1
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• total_train_batch_size: 4
• total_eval_batch_size: 4
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: inverse_sqrt
• lr_scheduler_warmup_steps: 100
• num_epochs: 3

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen	Nll Loss	Log Odds Ratio	Log Odds Chosen
1.3309	1.0	1259	1.4432	-0.0513	-0.0583	0.5468	0.0071	-1.1666	-1.0254	310.9833	338.2715	1.3964	-0.7034	0.2119
0.647	2.0	2518	1.4816	-0.0529	-0.0637	0.5899	0.0108	-1.2742	-1.0583	296.0398	324.3109	1.4304	-0.6778	0.3416
0.348	3.0	3777	1.7559	-0.0650	-0.0764	0.5971	0.0114	-1.5282	-1.3004	266.0260	295.6202	1.6941	-0.6992	0.3721

Framework versions

• Transformers 4.44.2
• Pytorch 2.4.0+cu121
• Datasets 3.0.0
• Tokenizers 0.19.1