zephyr-7b-dpo-full-prometheus_consistent-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: 0.5209
• Rewards/chosen: -1.1123
• Rewards/rejected: -2.3595
• Rewards/accuracies: 0.7672
• Rewards/margins: 1.2472
• Logps/rejected: -455.0307
• Logps/chosen: -386.8380
• Logits/rejected: 3.7050
• Logits/chosen: 2.1175

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
0.6663	0.1143	50	0.6532	-0.0060	-0.1654	0.6897	0.1594	-235.6116	-276.2017	-2.4378	-2.5412
0.6051	0.2286	100	0.6128	-0.8115	-1.2713	0.7112	0.4599	-346.2107	-356.7497	-2.3746	-2.4933
0.5375	0.3429	150	0.5487	-0.8314	-1.9411	0.7672	1.1097	-413.1859	-358.7466	1.3727	0.2031
0.5435	0.4571	200	0.5359	-0.9895	-2.0771	0.7543	1.0876	-426.7858	-374.5490	2.7076	1.3441
0.5433	0.5714	250	0.5290	-0.9003	-2.0663	0.7629	1.1660	-425.7063	-365.6371	2.4121	0.8588
0.5194	0.6857	300	0.5213	-0.9607	-2.2057	0.7716	1.2450	-439.6500	-371.6745	3.3117	1.7085
0.5325	0.8	350	0.5216	-1.1386	-2.3763	0.7629	1.2377	-456.7085	-389.4649	3.6708	2.0842
0.5483	0.9143	400	0.5209	-1.1123	-2.3595	0.7672	1.2472	-455.0307	-386.8380	3.7050	2.1175

Framework versions

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