# okto_version: "1.2"

# Teste 5: Flan-T5 com INFERENCE e BEHAVIOR
# Modelo: google/flan-t5-base
# Objetivo: Testar inferência com controle de comportamento

PROJECT "test_flan_t5_inference"
DESCRIPTION "Teste Flan-T5 com INFERENCE, BEHAVIOR e GUARD"

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: 3
  batch_size: 16
  learning_rate: 0.0001
  device: "auto"
}

BEHAVIOR {
  mode: "chat"
  personality: "friendly"
  verbosity: "medium"
  language: "pt-BR"
  avoid: ["violence", "hate", "politics"]
  fallback: "Como posso ajudar?"
  prompt_style: "User: {input}\nAssistant:"
}

GUARD {
  prevent {
    hallucination
    toxicity
    bias
    data_leak
  }
  detect_using: ["classifier", "regex", "rule_engine"]
  on_violation {
    REPLACE with_message: "Desculpe, essa solicitação não é permitida."
  }
}

INFERENCE {
  mode: "chat"
  format: "User: {input}\nAssistant:"
  exit_command: "/exit"
  params {
    max_length: 120
    temperature: 0.7
    top_p: 0.9
    beams: 2
    do_sample: true
  }
  CONTROL {
    IF confidence < 0.3 {
      RETRY
      LOG "Low confidence - retrying"
    }
    
    IF repetition > 3 {
      REGENERATE
      LOG "High repetition detected - regenerating"
    }
  }
}

MONITOR {
  metrics: ["loss", "val_loss", "accuracy", "confidence"]
  log_to: "logs/inference_test.log"
}

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