zephyr-7b-dpo-full-ultrabin-high-margin-3-epochs

This model is a fine-tuned version of alignment-handbook/zephyr-7b-sft-full on the HuggingFaceH4/ultrafeedback_binarized dataset. It achieves the following results on the evaluation set:

• Loss: 0.5894
• Rewards/chosen: -2.9914
• Rewards/rejected: -4.9379
• Rewards/accuracies: 0.7578
• Rewards/margins: 1.9464
• Logps/rejected: -756.4492
• Logps/chosen: -561.7738
• Logits/rejected: 3.7803
• Logits/chosen: 2.4993

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-07
• train_batch_size: 8
• eval_batch_size: 8
• seed: 55
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 2
• total_train_batch_size: 128
• total_eval_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.1
• 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
0.5611	0.3484	50	0.6184	-0.0925	-0.3566	0.6875	0.2640	-298.3184	-271.8824	-2.5030	-2.5443
0.3114	0.6969	100	0.5684	-1.2749	-2.2355	0.7188	0.9606	-486.2125	-390.1244	1.4736	0.8929
0.2115	1.0453	150	0.5424	-1.1893	-2.3764	0.7344	1.1871	-500.3030	-381.5569	1.7464	0.8516
0.1459	1.3937	200	0.5506	-1.5868	-2.9488	0.7383	1.3620	-557.5460	-421.3102	2.1181	1.2033
0.155	1.7422	250	0.5421	-1.7379	-3.1364	0.7422	1.3985	-576.3018	-436.4162	0.6639	-0.1257
0.0778	2.0906	300	0.5661	-2.2459	-3.9084	0.7578	1.6626	-653.5056	-487.2183	2.4478	1.3197
0.063	2.4390	350	0.5745	-2.4511	-4.2302	0.7461	1.7791	-685.6794	-507.7419	3.2009	2.0299
0.0546	2.7875	400	0.5884	-2.9614	-4.9020	0.7578	1.9406	-752.8591	-558.7693	3.7820	2.5008

Framework versions

• Transformers 4.44.0.dev0
• Pytorch 2.1.2
• Datasets 2.20.0
• Tokenizers 0.19.1