chieunq/dpo15

dpo

This model is a fine-tuned version of /raid/HUB_LLM/080225_vi_test_llama33_70b_instruct/checkpoint-2568/SLERP-1.5/ on the 60k dataset.

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: 9e-06
• train_batch_size: 2
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• total_train_batch_size: 16
• total_eval_batch_size: 64
• optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine_with_min_lr
• lr_scheduler_warmup_steps: 10
• num_epochs: 3.0

Training results

Framework versions

• PEFT 0.12.0
• Transformers 4.47.1
• Pytorch 2.5.1+cu124
• Datasets 3.1.0
• Tokenizers 0.21.0