New Features
Fine-tuning
- "Continue from model" now supports using a remote model from Hugging Face Hub.
- New advanced fine-tune arg:
lora_modules_to_save
- New advanced fine-tune arg:
8bit
,FP16
,BF16
,gradient_checkpointing
- Support passing additional LoraConfig and TrainingArguments.
- "Load training parameters from selected model" will now also load the prompt template and dataset.
- Model
trainalbe%
will be logged to wandb under configmodel.trainable%
,model.all_params
andmodel.trainable_params
.