zephyr-7b-dpo-full-debug-regression

This model is a fine-tuned version of HuggingFaceH4/mistral-7b-sft-beta on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.7240
• Rewards/chosen: -4.3843
• Rewards/rejected: -7.9101
• Rewards/accuracies: 0.7640
• Rewards/margins: 3.5258
• Logps/rejected: -311.4621
• Logps/chosen: -319.5667
• Logits/rejected: -2.4790
• Logits/chosen: -2.5088

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
• total_train_batch_size: 32
• 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
0.533	0.26	500	0.5084	-0.1902	-1.3680	0.7780	1.1778	-246.0413	-277.6251	-2.9319	-2.9487
0.4907	0.52	1000	0.5234	-0.3346	-1.8153	0.7620	1.4807	-250.5139	-279.0693	-2.8401	-2.8442
0.4388	0.77	1500	0.5202	-0.7856	-2.2720	0.7920	1.4864	-255.0812	-283.5798	-2.7420	-2.7444
0.0651	1.03	2000	0.5049	-1.0044	-2.8702	0.7860	1.8658	-261.0635	-285.7675	-2.7335	-2.7412
0.0887	1.29	2500	0.5946	-1.9888	-3.9256	0.7480	1.9368	-271.6175	-295.6113	-2.5940	-2.6173
0.0747	1.55	3000	0.5748	-1.9590	-4.0271	0.7560	2.0681	-272.6327	-295.3135	-2.4969	-2.5205
0.101	1.81	3500	0.5783	-1.9521	-4.1853	0.7680	2.2332	-274.2144	-295.2442	-2.5069	-2.5278
0.0195	2.07	4000	0.6253	-2.9322	-5.7633	0.7600	2.8310	-289.9938	-305.0455	-2.4935	-2.5158
0.0191	2.32	4500	0.7215	-4.2183	-7.6216	0.7620	3.4034	-308.5774	-317.9060	-2.4756	-2.5036
0.0105	2.58	5000	0.7341	-4.2607	-7.7440	0.7600	3.4833	-309.8016	-318.3306	-2.5156	-2.5437
0.0092	2.84	5500	0.7330	-4.3756	-7.9435	0.7600	3.5679	-311.7966	-319.4794	-2.4856	-2.5149

Framework versions

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