config.py

import torch
import os

# --- Paths ---
DATA_PATH = '/kaggle/input/synthesis-data/synthetic_transactions_samples_5000.csv'
TOKENIZER_PATH = './tokenizer_roberta/'
LABEL_ENCODERS_PATH = './label_encoders.pkl'
MODEL_SAVE_DIR = './saved_models/'
PREDICTIONS_SAVE_DIR = './predictions/'

# --- Data Columns ---
TEXT_COLUMN = "Sanction_Context"
LABEL_COLUMNS = [
    "Red_Flag_Reason",
    "Maker_Action",
    "Escalation_Level",
    "Risk_Category",
    "Risk_Drivers",
    "Investigation_Outcome"
]
METADATA_COLUMNS = []

# --- Model Hyperparameters ---
MAX_LEN = 128
BATCH_SIZE = 16
LEARNING_RATE = 2e-5
NUM_EPOCHS = 3
DROPOUT_RATE = 0.3

# --- Device Configuration ---
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# --- Model Names ---
BERT_MODEL_NAME = 'bert-base-uncased'
ROBERTA_MODEL_NAME = 'roberta-base'
DEBERTA_MODEL_NAME = 'microsoft/deberta-base'

# --- TF-IDF ---
TFIDF_MAX_FEATURES = 5000

# --- Strategy Definitions ---
FIELD_STRATEGIES = {
    "Maker_Action": {
        "loss": "focal_loss",
        "enhancements": ["action_templates", "context_prompt_tuning"]
    },
    "Risk_Category": {
        "enhancements": ["numerical_metadata", "transaction_patterns"]
    },
    "Escalation_Level": {
        "enhancements": ["class_balancing", "policy_keyword_patterns"]
    },
    "Investigation_Outcome": {
        "type": "classification_or_generation"
    }
}

# --- Ensure directories exist ---
os.makedirs(MODEL_SAVE_DIR, exist_ok=True)
os.makedirs(PREDICTIONS_SAVE_DIR, exist_ok=True)
os.makedirs(TOKENIZER_PATH, exist_ok=True)

# ✅ Set the active model for training/deployment
MODEL_NAME = ROBERTA_MODEL_NAME