phi-2-dpo-ultrafeedback-lora

This model is a fine-tuned version of lole25/phi-2-sft-ultrachat-lora on the HuggingFaceH4/ultrafeedback_binarized dataset. It achieves the following results on the evaluation set:

• Loss: 0.6537
• Rewards/chosen: -0.2570
• Rewards/rejected: -0.3767
• Rewards/accuracies: 0.6580
• Rewards/margins: 0.1196
• Logps/rejected: -269.1014
• Logps/chosen: -285.9487
• Logits/rejected: 0.7335
• Logits/chosen: 0.6309

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
• num_devices: 4
• gradient_accumulation_steps: 4
• total_train_batch_size: 64
• total_eval_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: 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
0.6929	0.21	100	0.6928	0.0002	-0.0010	0.5320	0.0012	-231.5360	-260.2240	0.9168	0.8145
0.6893	0.42	200	0.6891	-0.0038	-0.0134	0.6500	0.0096	-232.7742	-260.6225	0.9234	0.8205
0.6809	0.63	300	0.6810	-0.0312	-0.0611	0.6680	0.0299	-237.5431	-263.3647	0.9151	0.8092
0.6671	0.84	400	0.6723	-0.0854	-0.1408	0.6640	0.0553	-245.5124	-268.7867	0.8790	0.7713
0.6627	1.05	500	0.6645	-0.1494	-0.2293	0.6680	0.0799	-254.3704	-275.1849	0.8294	0.7217
0.6476	1.26	600	0.6591	-0.1979	-0.2968	0.6640	0.0989	-261.1124	-280.0337	0.7883	0.6828
0.6488	1.47	700	0.6559	-0.2310	-0.3414	0.6620	0.1104	-265.5783	-283.3440	0.7549	0.6511
0.6449	1.67	800	0.6542	-0.2518	-0.3695	0.6560	0.1177	-268.3814	-285.4226	0.7372	0.6347
0.6487	1.88	900	0.6539	-0.2571	-0.3764	0.6560	0.1193	-269.0724	-285.9532	0.7320	0.6299

Framework versions

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