zephyr-7b-dpo-full

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: 0.5058
• Rewards/chosen: -1.0287
• Rewards/rejected: -2.0159
• Rewards/accuracies: 0.7773
• Rewards/margins: 0.9873
• Logps/rejected: -464.2831
• Logps/chosen: -365.4620
• Logits/rejected: 0.4997
• Logits/chosen: -0.4859

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	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
0.5721	0.2092	100	0.5709	-0.7941	-1.3740	0.75	0.5798	-400.0867	-342.0089	-1.1226	-1.3180
0.5453	0.4184	200	0.5234	-1.0595	-1.8701	0.7773	0.8105	-449.6958	-368.5487	0.0691	-0.6851
0.4933	0.6276	300	0.5102	-1.0953	-2.0092	0.7695	0.9139	-463.6079	-372.1221	0.3510	-0.4567
0.4979	0.8368	400	0.5065	-0.9786	-1.9310	0.7734	0.9524	-455.7930	-360.4568	0.2947	-0.6442

Framework versions

• Transformers 4.40.2
• Pytorch 2.1.2+cu118
• Datasets 2.19.1
• Tokenizers 0.19.1