zephyr-7b-dpo-full

This model is a fine-tuned version of lewtun/zephyr-7b-sft-full on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.4949
• Rewards/chosen: -1.0109
• Rewards/rejected: -1.9949
• Rewards/accuracies: 0.7891
• Rewards/margins: 0.9840
• Logps/rejected: -458.5204
• Logps/chosen: -380.6149
• Logits/rejected: 1.7815
• Logits/chosen: 0.4467

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: 42
• 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
0.5594	0.21	100	-0.4568	0.0349	-345.0751	-391.4898	0.5459	0.7695	-0.6555	0.6691	-1.3246
0.5342	0.42	200	0.1644	1.0433	-360.3595	-414.7491	0.5221	0.8047	-0.8084	0.7489	-1.5572
0.5085	0.63	300	0.6383	1.9502	-383.1190	-455.6415	0.5045	0.7891	-1.0360	0.9302	-1.9662
0.4998	0.84	400	0.4907	1.8212	-384.5488	-461.4449	0.4953	0.7930	-1.0503	0.9739	-2.0242

Framework versions

• Transformers 4.36.2
• Pytorch 2.1.2+cu121
• Datasets 2.14.6
• Tokenizers 0.15.0