llama3.1-cpo_j-full-0911

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: 1.4373
• Rewards/chosen: -14.1493
• Rewards/rejected: -15.5710
• Rewards/accuracies: 0.6543
• Rewards/margins: 1.4217
• Logps/rejected: -155.7095
• Logps/chosen: -141.4926
• Logits/rejected: -0.1136
• Logits/chosen: -0.1476
• Nll Loss: 0.1725

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: 5

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
1.4367	0.9986	432	1.3926	-17.0679	-18.0962	0.6565	1.0283	-180.9624	-170.6792	-0.4080	-0.4373	0.3200
0.5472	1.9994	865	1.2973	-16.2909	-17.5852	0.6587	1.2944	-175.8523	-162.9086	-0.5434	-0.5688	0.2148
0.2244	2.9980	1297	1.3861	-15.7105	-17.2195	0.6565	1.5089	-172.1945	-157.1052	-0.3428	-0.3715	0.2034
0.1472	3.9988	1730	1.4029	-14.6462	-16.1385	0.6522	1.4923	-161.3849	-146.4623	-0.2701	-0.3029	0.1876
0.1143	4.9928	2160	1.4373	-14.1493	-15.5710	0.6543	1.4217	-155.7095	-141.4926	-0.1136	-0.1476	0.1725

Framework versions

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