zephyr-7b-dpo-reference

This model is a fine-tuned version of alignment-handbook/zephyr-7b-sft-full on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.0429
• Rewards/chosen: -0.5987
• Rewards/rejected: -10.2552
• Rewards/accuracies: 0.9741
• Rewards/margins: 9.6565
• Logps/rejected: -175.1052
• Logps/chosen: -304.8906
• Logits/rejected: -1.9643
• Logits/chosen: -2.1592

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
• total_train_batch_size: 32
• total_eval_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• 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
0.0811	0.29	500	0.0543	-0.5167	-8.9128	0.9720	8.3961	-161.6813	-304.0705	-2.0037	-2.1857
0.0362	0.57	1000	0.0483	-0.4980	-9.5824	0.9720	9.0844	-168.3771	-303.8834	-2.0113	-2.2030
0.0318	0.86	1500	0.0442	-0.8458	-10.5987	0.9720	9.7529	-178.5403	-307.3617	-1.9506	-2.1461

Framework versions

• Transformers 4.36.1
• Pytorch 2.0.1+cu117
• Datasets 2.16.1
• Tokenizers 0.15.0