zephyr-7b-dpo-full-hh

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

• Loss: 0.5341
• Rewards/chosen: -2.5112
• Rewards/rejected: -3.3276
• Rewards/accuracies: 0.7295
• Rewards/margins: 0.8164
• Logps/rejected: -485.8198
• Logps/chosen: -398.0228
• Logits/rejected: 2.4839
• Logits/chosen: 1.8909

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
0.6755	0.0796	100	0.6758	-0.1176	-0.1618	0.5830	0.0442	-169.2391	-158.6626	-2.4575	-2.4733
0.5965	0.1592	200	0.5934	-1.0644	-1.4934	0.6772	0.4290	-302.3988	-253.3484	-0.2911	-0.5108
0.5621	0.2388	300	0.5712	-1.3390	-1.8901	0.6875	0.5511	-342.0688	-280.8055	0.3164	-0.1607
0.551	0.3183	400	0.5651	-1.2110	-1.8230	0.7192	0.6120	-335.3575	-268.0063	-0.2086	-0.6012
0.5696	0.3979	500	0.5572	-1.8231	-2.3956	0.7229	0.5725	-392.6161	-329.2127	0.5151	0.0872
0.5504	0.4775	600	0.5508	-1.9233	-2.7091	0.7201	0.7858	-423.9663	-339.2298	1.2370	0.4867
0.5387	0.5571	700	0.5417	-2.3329	-3.1152	0.7211	0.7823	-464.5798	-380.1928	1.9107	1.1807
0.5119	0.6367	800	0.5416	-2.6721	-3.5027	0.7276	0.8306	-503.3281	-414.1180	3.4443	2.7209
0.564	0.7163	900	0.5385	-2.6361	-3.3606	0.7183	0.7245	-489.1202	-410.5185	2.5529	1.9952
0.5201	0.7959	1000	0.5347	-2.5021	-3.2845	0.7229	0.7824	-481.5121	-397.1160	2.4306	1.8888
0.5341	0.8754	1100	0.5346	-2.4898	-3.2841	0.7295	0.7943	-481.4664	-395.8830	2.3851	1.8147
0.5394	0.9550	1200	0.5341	-2.5107	-3.3276	0.7295	0.8168	-485.8161	-397.9764	2.4847	1.8912

Framework versions

• Transformers 4.41.2
• Pytorch 2.1.2
• Datasets 2.20.0
• Tokenizers 0.19.1