llama-3-8b-dpo-full

<!-- This model card has been generated automatically according to the information the Trainer had access to. You should probably proofread and complete it, then remove this comment. -->

llama-3-8b-dpo-full

This model is a fine-tuned version of meta-llama/Meta-Llama-3-8B-Instruct on the trl-lib/ultrafeedback_binarized dataset. It achieves the following results on the evaluation set:

• Loss: 0.6491
• Rewards/chosen: -0.1814
• Rewards/rejected: -0.2255
• Rewards/accuracies: 0.5625
• Rewards/margins: 0.0441
• Logps/rejected: -419.1795
• Logps/chosen: -335.9990
• Logits/rejected: -1.1373
• Logits/chosen: -1.0280

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: 3e-07
• train_batch_size: 2
• eval_batch_size: 4
• seed: 42
• distributed_type: multi-GPU
• num_devices: 32
• gradient_accumulation_steps: 8
• total_train_batch_size: 512
• total_eval_batch_size: 128
• 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

Framework versions

• Transformers 4.45.1
• Pytorch 2.4.0+cu121
• Datasets 3.0.0
• Tokenizers 0.20.0