zephyr-7b-dpo-full-ultrabin-high-margin

This model is a fine-tuned version of alignment-handbook/zephyr-7b-sft-full on the HuggingFaceH4/ultrafeedback_binarized dataset. It achieves the following results on the evaluation set:

• Loss: 0.5456
• Rewards/chosen: -0.6570
• Rewards/rejected: -1.6025
• Rewards/accuracies: 0.7734
• Rewards/margins: 0.9454
• Logps/rejected: -422.9099
• Logps/chosen: -328.3232
• Logits/rejected: 0.5127
• Logits/chosen: 0.0348

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: 55
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 2
• total_train_batch_size: 128
• total_eval_batch_size: 64
• 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	Logits/chosen	Logits/rejected	Logps/chosen	Logps/rejected	Validation Loss	Rewards/accuracies	Rewards/chosen	Rewards/margins	Rewards/rejected
0.3211	0.6969	100	-0.1294	0.2523	-336.0263	-427.7636	0.5572	0.75	-0.7341	0.9170	-1.6510

Framework versions

• Transformers 4.41.2
• Pytorch 2.1.2
• Datasets 2.20.0
• Tokenizers 0.19.1