diff --git a/training/configs/ornith-35b-lora.yaml b/training/configs/ornith-35b-lora.yaml index 1a6194f..cfec004 100644 --- a/training/configs/ornith-35b-lora.yaml +++ b/training/configs/ornith-35b-lora.yaml @@ -32,12 +32,12 @@ dataset: train_params: num_train_epochs: 3 per_device_train_batch_size: 1 - gradient_accumulation_steps: 8 + gradient_accumulation_steps: 16 # Increased to reduce per-step memory learning_rate: 0.0002 lr_scheduler_type: cosine weight_decay: 0.01 warmup_ratio: 0.03 # Will be converted to warmup_steps by TrainingArguments - max_seq_length: 1024 + max_seq_length: 512 # Reduced from 1024 to save memory logging_steps: 10 save_steps: 100 save_total_limit: 3