unsloth_finetuning_config.yaml

model_id: "unsloth/Phi-3-mini-4k-instruct" # see more here: https://huggingface.co/unsloth
output_model_id: "jhangmez//CHATPRG-v1.5-LLAMA-3-8B"
huggingface_dataset: "jhangmez/REGLAMENTO-GENERAL-E-INVESTIGACION-UNPRG"

system_message: "You are ChatPRG, a help chatbot for students and external individuals to learn about the general regulations and research regulations of the National University Pedro Ruiz Gallo, the alma mater university of great professionals from the region, Peru, and the world."

# script parameters
chat_template: "chatml" # Supports zephyr, chatml, mistral, llama, alpaca, vicuna, vicuna_old, unsloth
max_seq_len: 4096 # max sequence length for model and packing of the dataset
dtype: null  # Data type for the model
load_in_4bit: true  # Load model in 4-bit quantization
use_gradient_checkpointing: "unsloth"  # Use gradient checkpointing
random_state: 3407  # Random state for reproducibility
r: 16  # LoRA rank
lora_alpha: 16  # LoRA alpha value
lora_dropout: 0  # LoRA dropout rate
bias: "none"  # LoRA bias
use_rslora: false  # Use rank stabilized LoRA
loftq_config: null  # LoftQ configuration
learning_rate: 0.0002  # Learning rate for training
num_train_epochs: 3  # Number of training epochs
batch_size: 2  # Batch size per device
gradient_accumulation_steps: 4  # Gradient accumulation steps
warmup_steps: 5  # Warmup steps
lr_scheduler_type: "linear"
per_device_train_batch_size: 2
optim: "adamw_8bit"
logging_steps: 1
warmup_steps: 5

# others
output_dir: "outputs"
merged_16bit: True
merged_4bit: True
lora: True
f16: True
q4_k_m: True