zephyr-7b-dpo-full

This model is a fine-tuned version of Jessicahuang0018/zephyr-7b-sft-full on the HuggingFaceH4/ultrafeedback_binarized dataset. It achieves the following results on the evaluation set:

• Loss: 0.5239
• Rewards/chosen: -0.6658
• Rewards/rejected: -1.4726
• Rewards/accuracies: 0.7812
• Rewards/margins: 0.8069
• Logps/rejected: -411.1472
• Logps/chosen: -331.8063
• Logits/rejected: -1.0919
• Logits/chosen: -1.2155

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
0.6029	0.2092	100	0.6044	-0.1356	-0.4456	0.7031	0.3101	-308.4451	-278.7863	-2.6772	-2.7044
0.5641	0.4184	200	0.5478	-0.4828	-1.1391	0.7617	0.6563	-377.7956	-313.5140	-1.6117	-1.7057
0.5115	0.6276	300	0.5337	-0.5489	-1.2889	0.7656	0.7399	-392.7693	-320.1249	-1.2038	-1.3223
0.5248	0.8368	400	0.5252	-0.6177	-1.4052	0.7773	0.7875	-404.4023	-326.9973	-1.1062	-1.2274

Framework versions

• Transformers 4.45.2
• Pytorch 2.5.1+rocm6.2
• Datasets 3.5.0
• Tokenizers 0.20.3