zephyr-dpop-qlora-uf6k-5e-7

This model is a fine-tuned version of alignment-handbook/zephyr-7b-sft-full on the generation/UF6konly dataset. It achieves the following results on the evaluation set:

• Loss: 0.6916
• Positive Losses: 0.0091
• Dpo Losses: 0.6900
• Rewards/chosen: 0.0284
• Rewards/rejected: 0.0220
• Rewards/accuracies: 0.6570
• Rewards/margins: 0.0064
• Rewards/margins Max: 0.0305
• Rewards/margins Min: -0.0149
• Rewards/margins Std: 0.0150
• Logps/rejected: -256.3835
• Logps/chosen: -281.7567
• Logits/rejected: -2.7680
• Logits/chosen: -2.8065

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	Positive Losses	Dpo Losses	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
0.6925	0.3	100	0.6926	0.0030	0.6919	0.0172	0.0146	0.6330	0.0026	0.0140	-0.0079	0.0073	-257.1209	-282.8758	-2.7669	-2.8058
0.6928	0.61	200	0.6919	0.0086	0.6904	0.0259	0.0203	0.6470	0.0056	0.0265	-0.0130	0.0131	-256.5479	-282.0022	-2.7678	-2.8064
0.6916	0.91	300	0.6914	0.0070	0.6900	0.0283	0.0219	0.6630	0.0064	0.0305	-0.0148	0.0149	-256.3924	-281.7664	-2.7635	-2.8024

Framework versions

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