zephyr-7b-dpo-full

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: 0.5026
• Rewards/chosen: -1.0472
• Rewards/rejected: -1.9901
• Rewards/accuracies: 0.7619
• Rewards/margins: 0.9430
• Logps/rejected: -460.7913
• Logps/chosen: -388.8264
• Logits/rejected: 1.7625
• Logits/chosen: 1.0392

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: 615
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 4
• total_train_batch_size: 128
• total_eval_batch_size: 32
• 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
0.5736	0.21	100	0.5842	-0.3837	-0.8468	0.7242	0.4632	-346.4596	-322.4754	-2.3510	-2.4330
0.5116	0.42	200	0.5308	-0.9042	-1.7331	0.7520	0.8289	-435.0859	-374.5288	0.5012	0.0459
0.5027	0.63	300	0.5084	-0.8877	-1.7467	0.7639	0.8590	-436.4478	-372.8834	1.8224	1.1385
0.4823	0.84	400	0.5037	-1.1953	-2.1521	0.7619	0.9568	-476.9852	-403.6375	1.9978	1.3075

Framework versions

• Transformers 4.36.2
• Pytorch 2.1.2+cu121
• Datasets 2.14.6
• Tokenizers 0.15.0