gemma-7b-orpo-low-quality

This model is a fine-tuned version of google/gemma-7b on the silviasapora/low_quality_dpo7k dataset. It achieves the following results on the evaluation set:

• Loss: 1.5517
• Rewards/chosen: -0.0554
• Rewards/rejected: -0.0646
• Rewards/accuracies: 0.5612
• Rewards/margins: 0.0092
• Logps/rejected: -1.2920
• Logps/chosen: -1.1085
• Logits/rejected: 268.0282
• Logits/chosen: 297.1682
• Nll Loss: 1.4855
• Log Odds Ratio: -0.6970
• Log Odds Chosen: 0.2856

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: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 4
• total_train_batch_size: 32
• 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
1.436	0.9955	167	1.4679	-0.0508	-0.0571	0.5468	0.0063	-1.1420	-1.0158	288.9292	318.3812	1.4121	-0.6895	0.1983
1.1098	1.9970	335	1.4451	-0.0518	-0.0579	0.5468	0.0061	-1.1581	-1.0353	286.4312	315.0296	1.3839	-0.7228	0.2105
0.5921	2.9866	501	1.5517	-0.0554	-0.0646	0.5612	0.0092	-1.2920	-1.1085	268.0282	297.1682	1.4855	-0.6970	0.2856

Framework versions

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