zephyr-7b-dpo-full

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

• Loss: 0.5257
• Rewards/chosen: -0.6523
• Rewards/rejected: -1.4719
• Rewards/accuracies: 0.7695
• Rewards/margins: 0.8195
• Logps/rejected: -411.0257
• Logps/chosen: -329.0598
• Logits/rejected: 0.9901
• Logits/chosen: 0.7049

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: 42
• 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: 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
0.5992	0.2092	100	0.5956	-0.2932	-0.6564	0.7148	0.3632	-329.4821	-293.1491	-2.2402	-2.2843
0.57	0.4184	200	0.5591	-0.3908	-0.9608	0.7422	0.5700	-359.9165	-302.9073	-1.6390	-1.7197
0.5222	0.6276	300	0.5473	-0.4814	-1.1717	0.7461	0.6902	-381.0072	-311.9707	-1.3133	-1.4138
0.5332	0.8368	400	0.5284	-0.6175	-1.4117	0.7539	0.7941	-405.0050	-325.5808	0.5323	0.2839

Framework versions

• Transformers 4.40.2
• Pytorch 2.1.0a0+32f93b1
• Datasets 2.19.1
• Tokenizers 0.19.1