phi-2-gpo-renew2-b0.001-vllm-i1

This model is a fine-tuned version of DUAL-GPO/phi-2-gpo-renew2-b0.001-i0 on the HuggingFaceH4/ultrafeedback_binarized dataset. It achieves the following results on the evaluation set:

• Loss: 0.0674
• Rewards/chosen: 0.2141
• Rewards/rejected: 0.1823
• Rewards/accuracies: 0.5025
• Rewards/margins: 0.0317
• Logps/rejected: -1694.0739
• Logps/chosen: -2002.3224
• Logits/rejected: 0.1179
• Logits/chosen: 0.1215

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: 1

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
0.1015	0.16	100	0.0587	0.1574	0.1353	0.5010	0.0221	-1741.1520	-2059.0308	-0.0782	-0.0740
0.1063	0.32	200	0.0662	0.2082	0.1782	0.5035	0.0300	-1698.2368	-2008.2380	0.0246	0.0308
0.089	0.48	300	0.0674	0.2080	0.1779	0.5025	0.0302	-1698.5458	-2008.3832	0.1068	0.1094
0.0836	0.64	400	0.0678	0.2130	0.1815	0.5045	0.0315	-1694.8785	-2003.3719	0.0938	0.0992
0.0823	0.8	500	0.0675	0.2147	0.1828	0.5030	0.0319	-1693.5775	-2001.7003	0.1134	0.1173
0.095	0.96	600	0.0674	0.2143	0.1825	0.5025	0.0318	-1693.8877	-2002.1124	0.1193	0.1230

Framework versions

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