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: 0.4963
• Rewards/chosen: -1.3619
• Rewards/rejected: -2.3893
• Rewards/accuracies: 0.7619
• Rewards/margins: 1.0274
• Logps/rejected: -499.1414
• Logps/chosen: -418.1593
• Logits/rejected: 1.1001
• Logits/chosen: 0.2372

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: 4
• gradient_accumulation_steps: 2
• total_train_batch_size: 64
• total_eval_batch_size: 32
• 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
0.6267	0.1047	100	-2.4885	-2.4228	-298.8784	-298.6605	0.6219	0.7123	-0.1691	0.2154	-0.3845
0.5618	0.2093	200	-1.1643	-0.9025	-368.7673	-409.5043	0.5559	0.75	-0.8680	0.6250	-1.4930
0.5298	0.3140	300	0.1099	0.5848	-387.0007	-442.8027	0.5283	0.7679	-1.0503	0.7756	-1.8260
0.5585	0.4186	400	-1.2260	-0.6461	-368.9012	-419.9413	0.5200	0.7639	-0.8693	0.7280	-1.5973
0.5074	0.5233	500	0.1651	0.9655	-402.5899	-472.5687	0.5043	0.7698	-1.2062	0.9174	-2.1236
0.4678	0.6279	600	-0.1276	0.7528	-405.3750	-480.0774	0.4995	0.7698	-1.2341	0.9646	-2.1987
0.4767	0.7326	700	0.4974	-1.3438	-2.3581	0.7619	1.0142	-496.0146	-416.3518	1.1957	0.3567
0.475	0.8373	800	0.4971	-1.3985	-2.3889	0.7639	0.9904	-499.1005	-421.8184	1.0101	0.2102
0.4828	0.9419	900	0.4963	-1.3653	-2.3944	0.7639	1.0292	-499.6516	-418.4965	1.1202	0.2514

Framework versions

• Transformers 4.41.2
• Pytorch 2.3.1+cu121
• Datasets 2.19.2
• Tokenizers 0.19.1