zephyr-7b-dpo-full

This model is a fine-tuned version of alignment-handbook/zephyr-7b-sft-full on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.6641
• Rewards/chosen: -4.2025
• Rewards/rejected: -7.5064
• Rewards/accuracies: 0.7740
• Rewards/margins: 3.3038
• Logps/rejected: -318.9220
• Logps/chosen: -321.7840
• Logits/rejected: -2.6752
• Logits/chosen: -2.7102

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: 4
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 4
• total_train_batch_size: 128
• total_eval_batch_size: 16
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 3

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
0.5058	0.21	100	0.4911	-0.1160	-1.1492	0.7620	1.0332	-242.3301	-272.5491	-3.0425	-3.0547
0.4946	0.41	200	0.4825	-0.3097	-1.5239	0.7660	1.2143	-246.8446	-274.8822	-2.9173	-2.9312
0.4777	0.62	300	0.4710	-0.1682	-1.3949	0.7800	1.2266	-245.2894	-273.1778	-2.9493	-2.9706
0.4795	0.83	400	0.4613	-0.3814	-1.7739	0.7880	1.3925	-249.8568	-275.7468	-2.8663	-2.8864
0.1195	1.03	500	0.4680	-0.6715	-2.3344	0.7880	1.6629	-256.6091	-279.2414	-2.7880	-2.8173
0.1038	1.24	600	0.4890	-1.2315	-3.1990	0.7740	1.9675	-267.0262	-285.9885	-2.7944	-2.8235
0.1091	1.45	700	0.4969	-1.5189	-3.4689	0.7720	1.9499	-270.2775	-289.4514	-2.7692	-2.7947
0.1017	1.65	800	0.4938	-1.5010	-3.5680	0.7780	2.0670	-271.4719	-289.2352	-2.7970	-2.8233
0.0853	1.86	900	0.5150	-1.9917	-4.1892	0.7840	2.1975	-278.9565	-295.1473	-2.7667	-2.7959
0.0214	2.07	1000	0.5713	-2.9307	-5.6016	0.7740	2.6709	-295.9729	-306.4606	-2.7489	-2.7811
0.0169	2.27	1100	0.6226	-3.6197	-6.6310	0.7760	3.0112	-308.3753	-314.7625	-2.7096	-2.7449
0.0192	2.48	1200	0.6315	-3.8462	-6.9302	0.7760	3.0839	-311.9800	-317.4913	-2.6779	-2.7147
0.0157	2.68	1300	0.6585	-4.1276	-7.3886	0.7760	3.2610	-317.5031	-320.8812	-2.6830	-2.7171
0.018	2.89	1400	0.6696	-4.2900	-7.6252	0.7740	3.3352	-320.3536	-322.8378	-2.6703	-2.7057

Framework versions

• Transformers 4.35.0
• Pytorch 2.1.0+cu118
• Datasets 2.14.6
• Tokenizers 0.14.1