llama3.1-cpo-full

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: 1.6689
• Rewards/chosen: -15.0012
• Rewards/rejected: -15.8900
• Rewards/accuracies: 0.6336
• Rewards/margins: 0.8888
• Logps/rejected: -158.8998
• Logps/chosen: -150.0119
• Logits/rejected: -0.3381
• Logits/chosen: -0.3504
• Nll Loss: 0.4161

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: 8
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 8
• total_train_batch_size: 512
• total_eval_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 3

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
1.822	0.9238	100	1.7791	-14.6496	-15.4269	0.6034	0.7773	-154.2694	-146.4961	-0.4235	-0.4380	0.4058
1.5612	1.8476	200	1.6871	-15.1337	-15.9726	0.6379	0.8389	-159.7256	-151.3367	-0.3722	-0.3863	0.4197
1.3825	2.7714	300	1.6704	-15.1684	-16.0433	0.6293	0.8749	-160.4333	-151.6842	-0.3369	-0.3497	0.4209

Framework versions

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