training_arguments = SFTConfig(
output_dir=output_dir,
per_device_train_batch_size=16,
gradient_accumulation_steps=1,
optim="adamw_torch",
save_steps=100,
logging_steps=1,
learning_rate=1e-4,
bf16=True,
# max_grad_norm=0.3,
num_train_epochs=3,
save_strategy="epoch",
warmup_ratio=0.05,
group_by_length=True,
# lr_scheduler_type="cosine",
gradient_checkpointing=True,
gradient_checkpointing_kwargs = {"use_reentrant": True},
dataset_text_field="text",
max_seq_length=1024,
packing=False
# report_to="wandb",
)
baseline_model = 'zisisbatzos/2SFTs_llama3.2-3B'
Metrics:
emotion_accuracy = 0.36554621848739494
correct_format = 0.7058823529411765
