zephyr-7b-dpo-full-magpi-reward-scale-05

This model is a fine-tuned version of alignment-handbook/zephyr-7b-sft-full on an unknown dataset. It achieves the following results on the evaluation set:

• Loss: 0.0006
• Rewards/chosen: -1.9421
• Rewards/rejected: -75.0558
• Rewards/accuracies: 1.0
• Rewards/margins: 73.1137
• Logps/rejected: -8146.3696
• Logps/chosen: -561.1917
• Logits/rejected: 2.3223
• Logits/chosen: -0.9088

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
0.0079	0.1420	50	0.0052	-1.4109	-51.3621	0.9960	49.9512	-5776.9990	-508.0702	-2.4011	-2.8844
0.0031	0.2841	100	0.0012	-2.0608	-82.9205	1.0	80.8596	-8932.8350	-573.0657	0.0172	-2.4172
0.0016	0.4261	150	0.0008	-2.0420	-78.8194	1.0	76.7774	-8522.7256	-571.1802	1.8918	-2.1991
0.0015	0.5682	200	0.0007	-1.9757	-79.8216	1.0	77.8459	-8622.9443	-564.5500	2.5732	-0.8412
0.0016	0.7102	250	0.0008	-1.8298	-70.1341	1.0	68.3043	-7654.1978	-549.9620	2.2580	-1.2005
0.0008	0.8523	300	0.0006	-1.9079	-74.3261	1.0	72.4182	-8073.3999	-557.7685	2.2921	-0.9558
0.0029	0.9943	350	0.0006	-1.9421	-75.0558	1.0	73.1137	-8146.3696	-561.1917	2.3223	-0.9088

Framework versions

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