gemma-7b-dpo-full-mix2-beta-0.1

This model is a fine-tuned version of lewtun/gemma-7b-sft-full-deita-10k-v0 on the HuggingFaceH4/ultrafeedback_binarized and the HuggingFaceH4/orca_dpo_pairs datasets. It achieves the following results on the evaluation set:

• Loss: 0.4056
• Rewards/chosen: -0.3995
• Rewards/rejected: -3.5721
• Rewards/accuracies: 0.7926
• Rewards/margins: 3.1726
• Logps/rejected: -414.5198
• Logps/chosen: -392.3416
• Logits/rejected: 83.8425
• Logits/chosen: 83.1641

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
0.6146	0.18	100	0.4651	-2.3366	-4.2121	0.7527	1.8755	-420.9197	-411.7124	84.3991	81.8795
0.5464	0.35	200	0.4531	-0.7850	-3.1857	0.7899	2.4007	-410.6562	-396.1968	84.8764	82.9057
0.5841	0.53	300	0.4209	-1.5926	-4.2403	0.8085	2.6477	-421.2023	-404.2725	83.7612	81.9224
0.519	0.7	400	0.4162	-1.2384	-4.1774	0.7819	2.9390	-420.5732	-400.7308	85.8201	84.7816
0.5432	0.88	500	0.4134	-0.3763	-3.5060	0.8032	3.1296	-413.8586	-392.1099	83.6363	82.8991

Framework versions

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