gemma-7b-dpo-full-ultrafeedback-beta-0.01

This model is a fine-tuned version of lewtun/gemma-7b-sft-full-ultrachat-v0 on the HuggingFaceH4/ultrafeedback_binarized dataset. It achieves the following results on the evaluation set:

• Loss: 0.4698
• Rewards/chosen: -1.0027
• Rewards/rejected: -2.3339
• Rewards/accuracies: 0.7698
• Rewards/margins: 1.3312
• Logps/rejected: -1118.8601
• Logps/chosen: -1006.0907
• Logits/rejected: 90.6424
• Logits/chosen: 105.6680

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: 2
• eval_batch_size: 4
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 8
• total_train_batch_size: 128
• total_eval_batch_size: 32
• 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
0.552	0.21	100	0.5756	-2.8657	-3.5901	0.7460	0.7243	-1244.4771	-1192.3933	82.3244	96.5612
0.501	0.42	200	0.4914	-1.6427	-2.6660	0.7817	1.0233	-1152.0745	-1070.0895	91.1202	105.1467
0.4893	0.63	300	0.4810	-1.6604	-2.8398	0.7619	1.1794	-1169.4480	-1071.8550	87.4237	101.9799
0.4759	0.84	400	0.4718	-0.8508	-2.1538	0.7817	1.3030	-1100.8470	-990.8950	89.1600	104.0108

Framework versions

• Transformers 4.39.0.dev0
• Pytorch 2.1.2+cu121
• Datasets 2.14.6
• Tokenizers 0.15.1