llama3.1-cpo-full-0913

This model is a fine-tuned version of meta-llama/Meta-Llama-3.1-8B-Instruct on the princeton-nlp/llama3-ultrafeedback dataset. It achieves the following results on the evaluation set:

• Loss: 1.5934
• Rewards/chosen: -15.4936
• Rewards/rejected: -16.2190
• Rewards/accuracies: 0.6261
• Rewards/margins: 0.7255
• Logps/rejected: -162.1901
• Logps/chosen: -154.9355
• Logits/rejected: -0.4926
• Logits/chosen: -0.5160
• Nll Loss: 0.4228

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: 1e-06
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 8
• total_train_batch_size: 128
• total_eval_batch_size: 16
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• 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	Nll Loss
1.9304	0.2311	100	1.7873	-14.9945	-15.3576	0.5804	0.3632	-153.5762	-149.9445	-0.3649	-0.3854	0.4085
1.6908	0.4623	200	1.6702	-15.6437	-16.2439	0.5978	0.6002	-162.4385	-156.4369	-0.3777	-0.4014	0.4252
1.6317	0.6934	300	1.6162	-15.4682	-16.1519	0.6152	0.6837	-161.5185	-154.6818	-0.4753	-0.4948	0.4202
1.62	0.9246	400	1.5947	-15.5964	-16.3155	0.6261	0.7192	-163.1553	-155.9637	-0.4910	-0.5144	0.4262

Framework versions

• Transformers 4.44.2
• Pytorch 2.3.1
• Datasets 2.21.0
• Tokenizers 0.19.1