# Configuração didática QLoRA do roteador de sentimentos do EscutIA.
# Execute somente no Google Colab com GPU NVIDIA/CUDA após revisar os gates.

### model
model_name_or_path: Qwen/Qwen2.5-1.5B-Instruct
model_revision: 989aa79
trust_remote_code: true
quantization_method: bnb
quantization_bit: 4
quantization_type: nf4
double_quantization: true

### method
stage: sft
do_train: true
do_eval: true
finetuning_type: lora
lora_target: q_proj,k_proj,v_proj,o_proj
lora_rank: 8
lora_alpha: 16
lora_dropout: 0.05

### dataset
dataset_dir: ../dataset/dados/preparados
dataset: escutia_treino
eval_dataset: escutia_validacao
template: qwen
cutoff_len: 256

### output
output_dir: outputs/resultados/qlora_escutia_router
logging_strategy: steps
logging_steps: 10
save_strategy: epoch
save_total_limit: 2
save_only_model: false
overwrite_output_dir: false
report_to: none
run_name: qlora_escutia_router
plot_loss: true

### train
per_device_train_batch_size: 1
per_device_eval_batch_size: 1
gradient_accumulation_steps: 8
learning_rate: 0.0001
num_train_epochs: 2.0
lr_scheduler_type: cosine
warmup_ratio: 0.05
optim: adamw_torch
weight_decay: 0.01
max_grad_norm: 1.0
gradient_checkpointing: true
bf16: false
fp16: true
pure_bf16: false
dataloader_num_workers: 2
seed: 42

### eval
eval_strategy: epoch