zephyr-dpo-timedial

This model is a fine-tuned version of EllieS/zephyr-sft-timedial on the EllieS/timedial_dpo dataset. It achieves the following results on the evaluation set:

• Loss: 0.2236
• Rewards/chosen: 0.2987
• Rewards/rejected: -1.0958
• Rewards/accuracies: 1.0
• Rewards/margins: 1.3944
• Logps/rejected: -154.5925
• Logps/chosen: -0.6286
• Logits/rejected: -2.7419
• Logits/chosen: -2.7480

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-06
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• distributed_type: multi-GPU
• gradient_accumulation_steps: 2
• total_train_batch_size: 4
• 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
0.3995	0.35	100	0.3705	0.2856	-0.5202	1.0	0.8058	-97.0354	-1.9368	-2.7815	-2.7803
0.2236	0.69	200	0.2236	0.2987	-1.0958	1.0	1.3944	-154.5925	-0.6286	-2.7419	-2.7480

Framework versions

• PEFT 0.7.1
• Transformers 4.39.0.dev0
• Pytorch 2.1.2+cu121
• Datasets 2.14.6
• Tokenizers 0.15.2