zephyr-7b-dpo-qlora

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

• Loss: 0.5020
• Rewards/chosen: -2.1712
• Rewards/rejected: -2.9620
• Rewards/accuracies: 0.7054
• Rewards/margins: 0.7908
• Logps/rejected: -535.7245
• Logps/chosen: -475.6829
• Logits/rejected: -1.1754
• Logits/chosen: -1.2812

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-06
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• distributed_type: multi-GPU
• num_devices: 6
• gradient_accumulation_steps: 4
• total_train_batch_size: 96
• total_eval_batch_size: 24
• 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
0.6099	0.1570	100	0.5999	-0.4571	-0.6755	0.6369	0.2184	-307.0734	-304.2707	-2.1308	-2.2191
0.5468	0.3140	200	0.5354	-1.2063	-1.7025	0.6815	0.4962	-409.7740	-379.1950	-1.5157	-1.6078
0.5195	0.4710	300	0.5227	-1.5981	-2.2831	0.7083	0.6849	-467.8293	-418.3782	-1.3523	-1.4527
0.4895	0.6279	400	0.5142	-1.8555	-2.6654	0.6994	0.8099	-506.0622	-444.1171	-1.1070	-1.2180
0.4992	0.7849	500	0.5019	-2.3330	-3.1029	0.7054	0.7699	-549.8137	-491.8629	-1.1520	-1.2589
0.5001	0.9419	600	0.5021	-2.1712	-2.9608	0.7083	0.7896	-535.6057	-475.6837	-1.1768	-1.2825

Framework versions

• PEFT 0.13.2
• Transformers 4.45.2
• Pytorch 2.1.2
• Datasets 3.0.1
• Tokenizers 0.20.1