""" Run predictions with your fine-tuned scam classifier. Usage: python predict_scam.py --model_dir ./scam-classifier-model --text "You won a free prize!" """ import argparse import torch import os import subprocess from pathlib import Path import mlflow from dotenv import load_dotenv from transformers import AutoTokenizer, AutoModelForSequenceClassification REPO_ROOT = Path(__file__).resolve().parents[1] BRANCH_STAGE_MAP = { "feature/phase-2-audio-asr": "feature/phase-2-audio-asr", "feature/phase-3-serving-quantization": "feature/phase-3-serving-quantization", "feature/phase-1.5-ultimate-dataset": "model-modernbert-universal", "model-long-context": "model-modernbert-universal", "model-distilbert": "model-distilbert", "main": "main", } def detect_branch(default="main"): env_branch = os.getenv("DAGSHUB_BRANCH") or os.getenv("GIT_BRANCH") if env_branch: return env_branch.strip() result = subprocess.run( ["git", "branch", "--show-current"], cwd=REPO_ROOT, capture_output=True, text=True, check=False, ) branch = result.stdout.strip() return branch or default def stage_for_branch(branch): return BRANCH_STAGE_MAP.get(branch, branch.replace("/", "-") or "main") def get_device(): if torch.backends.mps.is_available(): return "mps" if torch.cuda.is_available(): return "cuda" return "cpu" def main(): load_dotenv() parser = argparse.ArgumentParser() parser.add_argument("--model_dir", type=str, default="mlflow", help="Path to local model, Hugging Face repo ID, or 'mlflow' to pull latest from DagsHub") parser.add_argument("--text", type=str, required=True) parser.add_argument("--registry_name", type=str, default=None, help="Name of the model in DagsHub MLflow Registry when using --model_dir=mlflow") args = parser.parse_args() branch = detect_branch() stage = stage_for_branch(branch) stage_slug = stage.replace("/", "-") registry_name = args.registry_name or f"ModernBERT-Scam-Classifier-{stage_slug}" # DagsHub Auth username = os.environ.get("MLFLOW_TRACKING_USERNAME") password = os.environ.get("MLFLOW_TRACKING_PASSWORD") repo_owner = os.environ.get("DAGSHUB_REPO_OWNER") repo_name = os.environ.get("DAGSHUB_REPO_NAME") if username and password: os.environ["DAGSHUB_USER"] = username os.environ["DAGSHUB_TOKEN"] = password import dagshub dagshub.auth.add_app_token(password) if repo_owner and repo_name: import dagshub dagshub.init(repo_name=repo_name, repo_owner=repo_owner, mlflow=True) # CRITICAL FIX: Force MLflow's internal boto3 client to route to DagsHub S3 instead of AWS os.environ["MLFLOW_S3_ENDPOINT_URL"] = f"https://dagshub.com/{repo_owner}/{repo_name}.s3" os.environ["AWS_ACCESS_KEY_ID"] = username os.environ["AWS_SECRET_ACCESS_KEY"] = password device = get_device() if args.model_dir == "mlflow": print("Fetching latest model from DagsHub MLflow registry...") mlflow.set_experiment(f"scam-detection/{stage}/inference") # Load directly from the Model Registry instead of searching for runs model_uri = f"models:/{registry_name}/latest" print(f"Loading model from registry: {model_uri}") try: pipeline = mlflow.transformers.load_model(model_uri, return_type="components") model = pipeline["model"] tokenizer = pipeline["tokenizer"] except Exception as e: print(f"ERROR: Failed to load model from registry ({e}). Please ensure you have trained and registered it.") return else: tokenizer = AutoTokenizer.from_pretrained(args.model_dir) model = AutoModelForSequenceClassification.from_pretrained(args.model_dir) model.to(device) model.eval() inputs = tokenizer(args.text, return_tensors="pt", truncation=True, max_length=8192).to(device) # Fix for some models crashing when token_type_ids are generated by the standard tokenizer if "token_type_ids" in inputs: del inputs["token_type_ids"] with torch.no_grad(): logits = model(**inputs).logits probs = torch.softmax(logits, dim=-1)[0] label = "SCAM" if probs[1] > probs[0] else "LEGIT" confidence = probs.max().item() print(f"Text: {args.text}") print(f"Prediction: {label} (confidence: {confidence:.2%})") print(f" Legit prob: {probs[0]:.2%} | Scam prob: {probs[1]:.2%}") if __name__ == "__main__": main()