zephyr-7b-dpo-full-prometheus_consistent-low-curriculum

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: 0.5227
• Rewards/chosen: -0.6988
• Rewards/rejected: -1.6296
• Rewards/accuracies: 0.7629
• Rewards/margins: 0.9308
• Logps/rejected: -382.0352
• Logps/chosen: -345.4830
• Logits/rejected: 0.7649
• Logits/chosen: -0.3664

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
0.6619	0.1143	50	0.6465	-0.0188	-0.1651	0.6940	0.1464	-235.5894	-277.4793	-2.4695	-2.5696
0.5843	0.2286	100	0.5720	-0.4443	-1.0204	0.7457	0.5761	-321.1169	-320.0323	-1.1914	-1.3850
0.5509	0.3429	150	0.5467	-0.4786	-1.2647	0.7328	0.7861	-345.5482	-323.4684	-0.6635	-1.3552
0.5275	0.4571	200	0.5396	-0.5220	-1.3807	0.7716	0.8587	-357.1487	-327.8021	-0.0895	-1.0377
0.5665	0.5714	250	0.5405	-1.0842	-1.9608	0.7629	0.8766	-415.1577	-384.0230	1.5254	0.2878
0.5202	0.6857	300	0.5275	-0.7424	-1.7347	0.7716	0.9923	-392.5496	-349.8434	0.9934	-0.2192
0.5261	0.8	350	0.5234	-0.7613	-1.7196	0.7586	0.9584	-391.0397	-351.7310	0.9008	-0.2697
0.5343	0.9143	400	0.5227	-0.6988	-1.6296	0.7629	0.9308	-382.0352	-345.4830	0.7649	-0.3664

Framework versions

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