lambda-llama-3-8b-dpo-test-orca

This model is a fine-tuned version of meta-llama/Meta-Llama-3.1-8B-Instruct on the HuggingFaceH4/ultrafeedback_binarized and the tanliboy/orca_dpo_pairs datasets. It achieves the following results on the evaluation set:

• Loss: 0.4795
• Rewards/chosen: -1.6860
• Rewards/rejected: -2.8132
• Rewards/accuracies: 0.7259
• Rewards/margins: 1.1272
• Logps/rejected: -645.5051
• Logps/chosen: -549.3651
• Logits/rejected: -2.6630
• Logits/chosen: -2.5985

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: 2e-07
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 4
• total_train_batch_size: 128
• total_eval_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• 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
0.6011	0.1744	100	0.5738	-0.8770	-1.2808	0.6988	0.4038	-492.2603	-468.4565	-2.4544	-2.4042
0.5447	0.3489	200	0.5242	-1.3236	-2.0879	0.7289	0.7644	-572.9752	-513.1177	-2.6319	-2.5732
0.5173	0.5233	300	0.5003	-1.6828	-2.6810	0.7259	0.9982	-632.2809	-549.0404	-2.6140	-2.5556
0.5144	0.6978	400	0.4851	-1.7107	-2.8135	0.7319	1.1028	-645.5279	-551.8306	-2.7027	-2.6365
0.5162	0.8722	500	0.4798	-1.7085	-2.8440	0.7259	1.1355	-648.5815	-551.6072	-2.6442	-2.5812

Framework versions

• Transformers 4.44.2
• Pytorch 2.4.0+cu121
• Datasets 2.19.1
• Tokenizers 0.19.1