examples/test-flan-t5-complete.okt

# okto_version: "1.2"

# Teste 4: Flan-T5 Completo - Todos os Blocos
# Modelo: google/flan-t5-base
# Objetivo: Testar todos os blocos avançados juntos

PROJECT "test_flan_t5_complete"
DESCRIPTION "Teste completo Flan-T5 com todos os blocos v1.2"

ENV {
  accelerator: "gpu"
  min_memory: "8GB"
  precision: "fp16"
  backend: "oktoseek"
  install_missing: true
}

DATASET {
  train: "dataset/train.jsonl"
  validation: "dataset/val.jsonl"
}

MODEL {
  base: "google/flan-t5-base"
  device: "auto"
}

TRAIN {
  epochs: 5
  batch_size: 16
  learning_rate: 0.0001
  device: "auto"
}

MONITOR {
  metrics: [
    "loss",
    "val_loss",
    "accuracy",
    "perplexity",
    "gpu_usage",
    "ram_usage",
    "throughput",
    "latency",
    "confidence"
  ]
  notify_if {
    loss > 2.0
    val_loss > 2.5
    gpu_usage > 90%
    ram_usage > 80%
  }
  log_to: "logs/training_complete.log"
}

CONTROL {
  on_step_end {
    LOG loss
  }
  
  on_epoch_end {
    SAVE model
    LOG "Epoch completed"
    
    IF loss > 1.5 {
      SET LR = 0.00005
      LOG "Loss still high after epoch - reducing LR"
    }
    
    IF accuracy > 0.9 {
      SAVE "best_model"
      LOG "High accuracy reached - saving best model"
    }
  }
  
  validate_every: 200
  
  IF loss > 2.0 {
    SET LR = 0.00005
    LOG "High loss detected"
  }
  
  IF val_loss > 2.5 {
    STOP_TRAINING
    LOG "Validation loss too high"
  }
  
  WHEN gpu_memory < 12GB {
    SET batch_size = 8
    LOG "Reducing batch size due to GPU pressure"
  }
  
  EVERY 1000 steps {
    SAVE checkpoint
    LOG "Periodic checkpoint"
  }
}

STABILITY {
  stop_if_nan: true
  stop_if_diverges: true
  min_improvement: 0.001
}

EXPORT {
  format: ["okm", "safetensors"]
  path: "export/"
}