Sync from GitHub (tests passed)

deep_learning/models/hub.py

@@ -20,6 +20,7 @@ _HF_TOKEN_ENV = "HF_TOKEN"
 _ARTIFACTS = [
     "best_tft_asro.ckpt",
     "pca_finbert.joblib",
+    "optuna_results.json",
 ]
 
 

deep_learning/training/hyperopt.py

@@ -204,7 +204,8 @@ def run_hyperopt(
     logger.info("Optuna best trial #%d: val_loss=%.6f", best.number, best.value)
     logger.info("Best params: %s", best.params)
 
-    results_path = Path(base_cfg.training.checkpoint_dir) / "optuna_results.json"
+    # Save alongside best_tft_asro.ckpt (tft/ root) so upload_tft_artifacts picks it up.
+    results_path = Path(base_cfg.training.best_model_path).parent / "optuna_results.json"
     results_path.parent.mkdir(parents=True, exist_ok=True)
     results_path.write_text(json.dumps({
         "best_trial": best.number,

deep_learning/training/trainer.py

@@ -286,7 +286,9 @@ def _apply_optuna_results(cfg: TFTASROConfig) -> TFTASROConfig:
     from dataclasses import replace
     from deep_learning.config import ASROConfig, TFTModelConfig, TrainingConfig
 
-    results_path = Path(cfg.training.checkpoint_dir) / "optuna_results.json"
+    # optuna_results.json is saved at tft/ root (alongside best_tft_asro.ckpt),
+    # not inside the checkpoints/ subdirectory.
+    results_path = Path(cfg.training.best_model_path).parent / "optuna_results.json"
     if not results_path.exists():
         return cfg