zephyr-7b-dpo-full-ultrabin-high-curriculum

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: -1.0085
• Rewards/rejected: -1.9531
• Rewards/accuracies: 0.7617
• Rewards/margins: 0.9446
• Logps/rejected: -457.9681
• Logps/chosen: -363.4786
• Logits/rejected: 2.0914
• Logits/chosen: 1.3032

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
0.6557	0.1046	50	0.6479	-0.0110	-0.1284	0.7070	0.1174	-275.5045	-263.7303	-2.6116	-2.6493
0.5639	0.2092	100	0.5652	-0.6096	-1.2098	0.7383	0.6002	-383.6375	-323.5883	-0.5618	-0.9012
0.5323	0.3138	150	0.5405	-0.5498	-1.2901	0.7617	0.7402	-391.6696	-317.6140	0.4792	-0.0900
0.536	0.4184	200	0.5354	-0.6382	-1.3831	0.7656	0.7449	-400.9734	-326.4470	0.6525	-0.0195
0.5163	0.5230	250	0.5185	-1.1124	-1.9604	0.7383	0.8480	-458.7008	-373.8662	2.4883	1.7620
0.5018	0.6276	300	0.5108	-0.9326	-1.8124	0.7578	0.8798	-443.9044	-355.8924	2.0905	1.3198
0.4999	0.7322	350	0.5094	-1.0356	-1.9491	0.7461	0.9135	-457.5764	-366.1917	2.0403	1.2353
0.4966	0.8368	400	0.5066	-0.9929	-1.9227	0.7578	0.9298	-454.9321	-361.9198	2.0226	1.2642
0.5198	0.9414	450	0.5057	-1.0106	-1.9562	0.7617	0.9455	-458.2784	-363.6942	2.1024	1.3120

Framework versions

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