examples/finetuning-llm.okt

PROJECT "FineTunedLLM"
DESCRIPTION "Fine-tuning a small LLM with checkpoint resume and custom callbacks"
VERSION "1.0"
AUTHOR "OktoSeek"

DATASET {
    train: "dataset/instruction_train.jsonl"
    validation: "dataset/instruction_val.jsonl"
    test: "dataset/instruction_test.jsonl"
    format: "instruction"
    type: "generation"
    language: "en"
}

MODEL {
    base: "oktoseek/base-llm-7b"
    architecture: "transformer"
    parameters: 7B
    context_window: 4096
    precision: "fp16"
}

TRAIN {
    epochs: 5
    batch_size: 4
    gradient_accumulation: 8
    learning_rate: 0.0001
    optimizer: "adamw"
    scheduler: "cosine_with_restarts"
    loss: "cross_entropy"
    device: "cuda"
    gpu: true
    mixed_precision: true
    early_stopping: true
    checkpoint_steps: 100
    checkpoint_path: "./checkpoints"
    weight_decay: 0.01
    gradient_clip: 1.0
    warmup_steps: 100
    save_strategy: "steps"
}

METRICS {
    loss
    perplexity
    bleu
    rouge_l
    token_efficiency
    response_coherence
}

VALIDATE {
    on_train: false
    on_validation: true
    frequency: 1
    save_best_model: true
    metric_to_monitor: "loss"
}

INFERENCE {
    max_tokens: 512
    temperature: 0.7
    top_p: 0.9
    top_k: 50
    repetition_penalty: 1.1
    stop_sequences: ["\n\nHuman:", "\n\nAssistant:"]
}

EXPORT {
    format: ["gguf", "onnx", "okm", "safetensors"]
    path: "export/"
    quantization: "int8"
    optimize_for: "speed"
}

DEPLOY {
    target: "api"
    endpoint: "http://localhost:9000/llm"
    requires_auth: true
    port: 9000
    max_concurrent_requests: 50
}

HOOKS {
    before_train: "scripts/preprocess_data.py"
    after_epoch: "scripts/custom_early_stop.py"
    on_checkpoint: "scripts/backup_checkpoint.sh"
}

LOGGING {
    save_logs: true
    metrics_file: "runs/finetuned-llm/metrics.json"
    training_file: "runs/finetuned-llm/training_logs.json"
    log_level: "info"
    log_every: 10
}