[Bug]: retrain.py corrupts the model - duplicate __main__, train/test leakage, and a pre-training reload trigger

[pavsoss]

Bug Description

backend/retrain.py has two if __name__ == "__main__": blocks that both run
top-to-bottom, so trigger_model_reload() fires before main() retrains,
reloading the old model. The training logic is also broken:

• Import-time side effect: df = pd.read_csv('dataset.csv') runs at module
  import and crashes if the file is absent.
• Train/test leakage + double-fit: the vectorizer is fit on the train split,
  then immediately re-fit on the entire df, and model.fit(X, y) retrains on top
  of the split-trained model using raw string labels from the global df, not the
  label_encoder-encoded combined labels used elsewhere.
• Inconsistent persistence: it joblib.dumps the "final" artifacts and then
  pickle.dumps different objects to the same paths, so what lands on disk
  depends on line order.

Location: backend/retrain.py.

Steps to Reproduce

1. cd backend && python retrain.py.
2. Observe the reload trigger firing before training, and inspect the held-out
   accuracy vs. the actually-persisted model.

Expected Behavior

A deterministic pipeline: no import-time I/O; fit once on the train split for the
held-out metric; refit once on full combined data with consistent label encoding;
a single serialization path with atomic writes; reload triggered only after a
successful save.

Actual Behavior

Reload runs before training, the persisted model is the product of leaked/double
fits with mismatched labels, and artifacts are serialized twice inconsistently.

Screenshots

No response

Additional Information

Add tests/test_retrain.py using a tiny synthetic dataset + feedback CSV to assert
no leakage, label consistency, and artifact round-trip.

[github-actions]

👋 Welcome back! Thanks for your continued contribution. Your issue/PR is received and waiting for maintainer review.

[github-actions]

The issue describes multiple problems with the retrain.py script, including redundant __main__ blocks that cause pre-training model reloads, broken training logic leading to train/test leakage and inconsistent label encoding, and conflicting artifact serialization methods (joblib vs. pickle). These flaws result in corrupted models, unpredictable persistence, and inaccurate metrics during the training process. The script requires fixes to ensure deterministic operation, proper data handling, and consistent serialization behavior.