zephyr-7b-dpo-full-gpt_consistent-reward-scale-01

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: 0.5112
• Rewards/chosen: -2.1769
• Rewards/rejected: -3.5598
• Rewards/accuracies: 0.75
• Rewards/margins: 1.3829
• Logps/rejected: -602.5067
• Logps/chosen: -502.7829
• Logits/rejected: 3.1646
• Logits/chosen: 1.7737

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
0.6717	0.1147	50	0.6610	-0.0092	-0.1379	0.7069	0.1287	-260.3159	-286.0137	-2.4941	-2.5738
0.5954	0.2294	100	0.5793	-1.0175	-1.6882	0.6724	0.6706	-415.3387	-386.8440	0.2130	-0.2228
0.5532	0.3440	150	0.5526	-1.2662	-2.2135	0.7069	0.9473	-467.8689	-411.7078	1.5934	0.5662
0.5464	0.4587	200	0.5261	-1.5551	-2.7684	0.7371	1.2134	-523.3669	-440.5982	3.1764	1.9612
0.5314	0.5734	250	0.5204	-1.7538	-2.9498	0.7155	1.1960	-541.4977	-460.4667	2.5752	1.2547
0.5372	0.6881	300	0.5156	-1.9506	-3.1990	0.7543	1.2484	-566.4211	-480.1506	2.6898	1.2744
0.5302	0.8028	350	0.5109	-2.0192	-3.3382	0.7586	1.3190	-580.3456	-487.0094	2.8531	1.4191
0.5215	0.9174	400	0.5112	-2.1769	-3.5598	0.75	1.3829	-602.5067	-502.7829	3.1646	1.7737

Framework versions

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