• Model creator: https://huggingface.co/Minbyul/
• Original model: https://huggingface.co/Minbyul/meditron-7b-dpo-full-sft-wo-kqa_golden/

Original model description:

• alignment-handbook
• trl
• dpo
• generated_from_trainer
• trl
• dpo
• generated_from_trainer datasets:
• HuggingFaceH4/ultrafeedback_binarized model-index:
• name: meditron-7b-dpo-full-sft-wo-kqa_golden results: []

meditron-7b-dpo-full-sft-wo-kqa_golden

• Loss: 0.6415
• Rewards/chosen: -0.0188
• Rewards/rejected: -0.1421
• Rewards/accuracies: 0.7750
• Rewards/margins: 0.1233
• Logps/rejected: -693.1373
• Logps/chosen: -157.5604
• Logits/rejected: -1.0479
• Logits/chosen: -1.3209

Model description

Intended uses & limitations

Training and evaluation data

Training procedure

Training hyperparameters

• learning_rate: 5e-07
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 2
• total_train_batch_size: 64
• 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

Framework versions

• Transformers 4.39.0.dev0
• Pytorch 2.1.2
• Datasets 2.14.6
• Tokenizers 0.15.2