zephyr-7b-dpo-lora

This model is a fine-tuned version of allenai/tulu-2-7b on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.6551
• Rewards/chosen: -0.0635
• Rewards/rejected: -0.1612
• Rewards/accuracies: 0.6905
• Rewards/margins: 0.0978
• Logps/rejected: -241.7965
• Logps/chosen: -271.5237
• Logits/rejected: -1.4412
• Logits/chosen: -1.2746

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: 2
• eval_batch_size: 4
• seed: 42
• distributed_type: multi-GPU
• num_devices: 6
• gradient_accumulation_steps: 8
• total_train_batch_size: 96
• total_eval_batch_size: 24
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 3

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
0.6812	1.0	645	0.6832	-0.0153	-0.0425	0.6190	0.0271	-240.6086	-271.0424	-1.4556	-1.2853
0.667	2.0	1291	0.6636	-0.0473	-0.1169	0.6518	0.0695	-241.3527	-271.3625	-1.4457	-1.2786
0.6518	3.0	1935	0.6551	-0.0635	-0.1612	0.6905	0.0978	-241.7965	-271.5237	-1.4412	-1.2746

Framework versions

• Transformers 4.35.0
• Pytorch 2.1.1+cu121
• Datasets 2.14.6
• Tokenizers 0.14.1