llama3.1-cpo-full-0912

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: 1.5985
• Rewards/chosen: -15.4365
• Rewards/rejected: -16.1367
• Rewards/accuracies: 0.6239
• Rewards/margins: 0.7002
• Logps/rejected: -161.3668
• Logps/chosen: -154.3647
• Logits/rejected: -0.3853
• Logits/chosen: -0.4112
• Nll Loss: 0.4210

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
1.9362	0.2311	100	1.7930	-14.9339	-15.2848	0.5761	0.3508	-152.8475	-149.3394	-0.4123	-0.4378	0.4067
1.7019	0.4623	200	1.6786	-15.4303	-16.0131	0.6087	0.5828	-160.1311	-154.3027	-0.3358	-0.3580	0.4193
1.6388	0.6934	300	1.6233	-15.5465	-16.2127	0.6130	0.6662	-162.1269	-155.4650	-0.3582	-0.3828	0.4230
1.632	0.9246	400	1.6007	-15.6505	-16.3448	0.6370	0.6943	-163.4479	-156.5048	-0.3811	-0.4072	0.4277

Framework versions

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