reward-judge_iter-dpo-genRM_pilot-exp_iter1

This model is a fine-tuned version of RyanYr/reward-judge_SFT-genRM_pilot-exp on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.2657
• Rewards/chosen: -1.3929
• Rewards/rejected: -3.4254
• Rewards/accuracies: 0.8400
• Rewards/margins: 2.0324
• Logps/rejected: -188.0495
• Logps/chosen: -189.5504
• Logits/rejected: -0.7024
• Logits/chosen: -0.6169

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: 2e-07
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 128
• total_train_batch_size: 512
• total_eval_batch_size: 4
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 100
• num_epochs: 2

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
0.5701	0.9370	75	0.5377	0.1101	-0.2637	0.8400	0.3738	-156.4332	-174.5201	-0.6949	-0.6098
0.2655	1.8739	150	0.2657	-1.3929	-3.4254	0.8400	2.0324	-188.0495	-189.5504	-0.7024	-0.6169

Framework versions

• Transformers 4.43.4
• Pytorch 2.1.2+cu121
• Datasets 2.14.6
• Tokenizers 0.19.1