zephyr-7b-dpo-full-prometheus-reward-scale-1

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: 0.4890
• Rewards/chosen: -1.5856
• Rewards/rejected: -2.7832
• Rewards/accuracies: 0.7371
• Rewards/margins: 1.1976
• Logps/rejected: -526.5942
• Logps/chosen: -418.5211
• Logits/rejected: 3.6380
• Logits/chosen: 2.5080

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
0.6615	0.1143	50	0.6475	-0.0448	-0.1732	0.6595	0.1284	-265.5944	-264.4379	-2.5754	-2.6221
0.5507	0.2286	100	0.5504	-0.7596	-1.5622	0.6897	0.8026	-404.4943	-335.9164	0.2567	-0.1843
0.521	0.3429	150	0.5234	-0.9406	-1.8921	0.7198	0.9516	-437.4890	-354.0176	1.8188	1.1097
0.5162	0.4571	200	0.5108	-1.2309	-2.2537	0.7328	1.0228	-473.6453	-383.0527	2.2567	1.2299
0.5186	0.5714	250	0.5027	-1.3697	-2.4832	0.7026	1.1135	-496.5983	-396.9323	2.8322	1.5762
0.4883	0.6857	300	0.4936	-1.5404	-2.7021	0.7371	1.1617	-518.4875	-413.9998	3.3276	2.1394
0.4714	0.8	350	0.4907	-1.5101	-2.6852	0.7457	1.1752	-516.7975	-410.9666	3.5416	2.3945
0.4977	0.9143	400	0.4890	-1.5856	-2.7832	0.7371	1.1976	-526.5942	-418.5211	3.6380	2.5080

Framework versions

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