phi-2-gpo-test-longest-iter-v1-4

This model is a fine-tuned version of DUAL-GPO/phi-2-gpo-test-longest-iter-v1-3 on the HuggingFaceH4/ultrafeedback_binarized dataset. It achieves the following results on the evaluation set:

• Loss: 0.0015
• Rewards/chosen: 0.0002
• Rewards/rejected: -0.0007
• Rewards/accuracies: 0.5360
• Rewards/margins: 0.0010
• Logps/rejected: -278.6109
• Logps/chosen: -306.2527
• Logits/rejected: 0.0934
• Logits/chosen: -0.0054

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: 4
• eval_batch_size: 4
• seed: 42
• distributed_type: multi-GPU
• gradient_accumulation_steps: 4
• total_train_batch_size: 16
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 4

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
0.001	1.6	100	0.0015	0.0000	-0.0003	0.5040	0.0004	-278.5704	-306.2733	0.0926	-0.0051
0.001	3.2	200	0.0015	-0.0000	-0.0004	0.4885	0.0004	-278.5775	-306.2765	0.0977	-0.0020

Framework versions

• PEFT 0.7.1
• Transformers 4.36.2
• Pytorch 2.2.1+cu121
• Datasets 2.14.6
• Tokenizers 0.15.2