PROJECT "QAEmbeddings"
DESCRIPTION "Question-Answering system with embeddings and similarity search"
VERSION "1.0"
AUTHOR "OktoSeek"

DATASET {
    train: "dataset/qa_train.jsonl"
    validation: "dataset/qa_val.jsonl"
    format: "qa"
    type: "qa"
    language: "en"
}

MODEL {
    base: "oktoseek/qa-encoder"
    architecture: "bert"
    parameters: 110M
    context_window: 512
    precision: "fp16"
}

TRAIN {
    epochs: 10
    batch_size: 16
    learning_rate: 0.00005
    optimizer: "adamw"
    scheduler: "linear"
    loss: "cross_entropy"
    device: "cuda"
    gpu: true
    mixed_precision: true
    early_stopping: true
    checkpoint_steps: 200
    weight_decay: 0.01
    gradient_clip: 1.0
    warmup_steps: 500
}

METRICS {
    accuracy
    f1
    f1_macro
    cosine_similarity
    custom "retrieval_accuracy"
}

VALIDATE {
    on_validation: true
    frequency: 1
    save_best_model: true
    metric_to_monitor: "f1"
}

INFERENCE {
    max_tokens: 256
    temperature: 0.3
    top_p: 0.95
    top_k: 20
}

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

DEPLOY {
    target: "api"
    endpoint: "http://localhost:9000/qa"
    requires_auth: true
    port: 9000
    max_concurrent_requests: 200
}

LOGGING {
    save_logs: true
    metrics_file: "runs/qa-embeddings/metrics.json"
    training_file: "runs/qa-embeddings/training_logs.json"
    log_level: "info"
    log_every: 20
}