zephyr-dpo-qlora-gpt4-5e-7

This model is a fine-tuned version of alignment-handbook/zephyr-7b-sft-full on the generation/GPT4 dataset. It achieves the following results on the evaluation set:

• Loss: 0.6861
• Rewards/chosen: -0.0552
• Rewards/rejected: -0.0760
• Rewards/accuracies: 0.5910
• Rewards/margins: 0.0208
• Rewards/margins Max: 0.1489
• Rewards/margins Min: -0.0942
• Rewards/margins Std: 0.0802
• Logps/rejected: -266.1821
• Logps/chosen: -290.1167
• Logits/rejected: -2.7232
• Logits/chosen: -2.7609

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: 4
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 2
• gradient_accumulation_steps: 2
• total_train_batch_size: 16
• total_eval_batch_size: 16
• 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	Rewards/margins Max	Rewards/margins Min	Rewards/margins Std	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
0.6666	0.28	100	0.6907	-0.0061	-0.0128	0.5780	0.0067	0.0489	-0.0295	0.0259	-259.8633	-285.2067	-2.7596	-2.7978
0.6251	0.56	200	0.6876	-0.0360	-0.0522	0.5870	0.0162	0.1195	-0.0752	0.0641	-263.7989	-288.1948	-2.7319	-2.7700
0.6029	0.85	300	0.6862	-0.0539	-0.0744	0.5870	0.0206	0.1479	-0.0932	0.0795	-266.0218	-289.9789	-2.7191	-2.7573

Framework versions

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