llama3.1-cpo_j-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: 1.4404
• Rewards/chosen: -16.0845
• Rewards/rejected: -16.8741
• Rewards/accuracies: 0.6326
• Rewards/margins: 0.7896
• Logps/rejected: -168.7413
• Logps/chosen: -160.8449
• Logits/rejected: -0.3487
• Logits/chosen: -0.3704
• Nll Loss: 0.2790

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
1.7833	0.2311	100	1.6389	-15.3939	-15.7792	0.5783	0.3853	-157.7921	-153.9390	-0.3065	-0.3333	0.2678
1.5321	0.4623	200	1.5242	-15.8988	-16.5121	0.5978	0.6132	-165.1206	-158.9884	-0.4244	-0.4423	0.2764
1.4722	0.6934	300	1.4633	-16.0803	-16.8141	0.6217	0.7338	-168.1411	-160.8031	-0.3641	-0.3856	0.2790
1.4589	0.9246	400	1.4447	-16.0215	-16.8041	0.6261	0.7826	-168.0413	-160.2150	-0.3389	-0.3606	0.2798

Framework versions

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