gemma-7b-borpo-low-quality-v5

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: 2.0035
• Rewards/chosen: -0.6455
• Rewards/rejected: -0.7620
• Rewards/accuracies: 0.6115
• Rewards/margins: 0.1164
• Logps/rejected: -1.5240
• Logps/chosen: -1.2911
• Logits/rejected: 259.2041
• Logits/chosen: 292.7468
• Nll Loss: 1.6440
• Log Odds Ratio: -0.6769
• Log Odds Chosen: 0.3357

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: 7.5e-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
1.8398	0.9955	167	1.8982	-0.5762	-0.6715	0.5180	0.0953	-1.3431	-1.1525	302.9095	331.1552	1.5454	-0.6749	0.2542
1.299	1.9970	335	1.8186	-0.5507	-0.6510	0.5396	0.1003	-1.3021	-1.1014	282.3183	313.4974	1.4682	-0.6666	0.3074
0.6379	2.9866	501	2.0035	-0.6455	-0.7620	0.6115	0.1164	-1.5240	-1.2911	259.2041	292.7468	1.6440	-0.6769	0.3357

Framework versions

• Transformers 4.45.1
• Pytorch 2.4.1+cu121
• Datasets 3.0.1
• Tokenizers 0.20.0