zephyr-7b-dpo-full-repnew

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: 0.5036
• Rewards/chosen: -1.1057
• Rewards/rejected: -2.0459
• Rewards/accuracies: 0.7698
• Rewards/margins: 0.9402
• Logps/rejected: -466.3643
• Logps/chosen: -394.6778
• Logits/rejected: 0.8720
• Logits/chosen: 0.0385

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: 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	Logits/chosen	Logits/rejected	Logps/chosen	Logps/rejected	Validation Loss	Rewards/accuracies	Rewards/chosen	Rewards/margins	Rewards/rejected
0.5666	0.21	100	-1.6453	-1.5540	-378.9401	-411.0335	0.5780	0.7282	-0.9484	0.5442	-1.4926
0.5107	0.42	200	-0.1291	0.3999	-386.8341	-445.9254	0.5233	0.7480	-1.0273	0.8142	-1.8415
0.5036	0.63	300	-0.0425	0.7446	-387.1995	-449.6839	0.5109	0.7599	-1.0310	0.8481	-1.8791
0.485	0.84	400	0.0635	0.9022	-397.1799	-468.7184	0.5047	0.7639	-1.1308	0.9387	-2.0694

Framework versions

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