zephyr-7b-gpo-iter2

This model is a fine-tuned version of DUAL-GPO/zephyr-7b-gpo-iter1 on the HuggingFaceH4/ultrafeedback_binarized dataset. It achieves the following results on the evaluation set:

• Loss: 0.0114
• Rewards/chosen: -0.0874
• Rewards/rejected: -0.0645
• Rewards/accuracies: 0.3940
• Rewards/margins: -0.0229
• Logps/rejected: -264.6114
• Logps/chosen: -288.2511
• Logits/rejected: -2.1907
• Logits/chosen: -2.3882

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: 2
• seed: 42
• distributed_type: multi-GPU
• gradient_accumulation_steps: 2
• total_train_batch_size: 2
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 2

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
0.0012	0.3	100	0.0016	-0.0164	-0.0160	0.5035	-0.0005	-259.7555	-281.1500	-2.1644	-2.3583
0.0011	0.61	200	0.0018	-0.0088	-0.0077	0.4815	-0.0011	-258.9317	-280.3858	-2.1837	-2.3781
0.0015	0.91	300	0.0019	-0.0167	-0.0149	0.4805	-0.0017	-259.6521	-281.1740	-2.1796	-2.3740
0.0397	1.22	400	0.0074	-0.0779	-0.0627	0.4160	-0.0151	-264.4323	-287.2935	-2.1632	-2.3568
0.0305	1.52	500	0.0117	-0.0898	-0.0668	0.3945	-0.0230	-264.8388	-288.4842	-2.1902	-2.3875
0.0366	1.82	600	0.0115	-0.0876	-0.0647	0.4000	-0.0230	-264.6301	-288.2723	-2.1900	-2.3873

Framework versions

• PEFT 0.7.1
• Transformers 4.36.2
• Pytorch 2.1.2+cu118
• Datasets 2.14.6
• Tokenizers 0.15.2