zephyr-7b-dpo-full-ultrabin-reward-scale-1

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: 0.5232
• Rewards/chosen: -1.5749
• Rewards/rejected: -2.8516
• Rewards/accuracies: 0.7656
• Rewards/margins: 1.2768
• Logps/rejected: -547.8262
• Logps/chosen: -420.1173
• Logits/rejected: 1.4432
• Logits/chosen: 0.5663

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	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
0.6644	0.1046	50	0.6488	0.0260	-0.0889	0.7109	0.1149	-271.5553	-260.0293	-2.5794	-2.6162
0.5487	0.2092	100	0.5902	-1.2155	-1.9393	0.7227	0.7239	-456.5951	-384.1769	0.4524	0.1576
0.5256	0.3138	150	0.5479	-1.0848	-2.0806	0.7344	0.9958	-470.7209	-371.1125	1.5020	0.9718
0.5215	0.4184	200	0.5366	-1.2109	-2.2989	0.7656	1.0880	-492.5526	-383.7176	1.7166	1.0342
0.5123	0.5230	250	0.5273	-1.0432	-2.1107	0.7617	1.0675	-473.7350	-366.9494	0.9023	0.2939
0.4963	0.6276	300	0.5263	-1.4844	-2.7207	0.7734	1.2363	-534.7312	-411.0707	1.6647	0.8936
0.4915	0.7322	350	0.5202	-1.5546	-2.7469	0.7773	1.1924	-537.3561	-418.0878	1.2880	0.4607
0.5018	0.8368	400	0.5239	-1.6225	-2.8974	0.7656	1.2750	-552.4055	-424.8771	1.5572	0.6801
0.497	0.9414	450	0.5231	-1.5820	-2.8565	0.7695	1.2746	-548.3168	-420.8284	1.4463	0.5693

Framework versions

• Transformers 4.44.0.dev0
• Pytorch 2.1.2
• Datasets 2.20.0
• Tokenizers 0.19.1