fix: make forecast window text dynamic based on actual horizon selected

agents/nodes.py

@@ -452,7 +452,8 @@ def strategy_formulator_node(state: GridOpsState) -> dict:
     # Get the LLM summary
     summary_rationale = mandate.get('summary_rationale', '')
 
-    static_intro = "The GridOps AI pipeline has analyzed the current 14-day forecast window. Below is the operational assessment of the predictive models."
+    horizon = state.get('forecast_horizon', 14)
+    static_intro = f"The GridOps AI pipeline has analyzed the current {horizon}-day forecast window. Below is the operational assessment of the predictive models."
 
     mandate_narrative = (
         f"{static_intro}\n\n"
@@ -552,7 +553,8 @@ def conservative_advisory_node(state: GridOpsState) -> dict:
     
     summary_rationale = mandate.get('summary_rationale', '')
 
-    static_intro = "The GridOps AI pipeline has analyzed the current 14-day forecast window. Below is the operational assessment of the predictive models."
+    horizon = state.get('forecast_horizon', 14)
+    static_intro = f"The GridOps AI pipeline has analyzed the current {horizon}-day forecast window. Below is the operational assessment of the predictive models."
 
     mandate_narrative = (
         f"{static_intro}\n\n"

scripts/train_chronos_mac.py

@@ -57,13 +57,14 @@ def main():
                 "batch_size": 2,              # Inference batch size
                 "fine_tune_batch_size": 1,    # ABSOLUTE MINIMUM batch to fit 200M model in 8GB RAM
                 # --- Anti-Overfitting Training settings ---
-                "fine_tune_steps": 150,       # Early stopping (150 steps instead of 500) to prevent memorizing noise
-                "fine_tune_lr": 3e-5,         # Lower learning rate to preserve the base model's generalized knowledge
+                "fine_tune_steps": 5000,      # Train for 5000 steps to properly learn the PJM patterns
+                "fine_tune_lr": 1e-5,         # Smaller learning rate for a longer training run
                 # Force non-fused optimizer (fused AdamW requires CUDA, not MPS)
                 "fine_tune_trainer_kwargs": {
                     "optim": "adamw_torch",
+                    "gradient_accumulation_steps": 8,  # Simulates a batch size of 8 while only using RAM for 1
                     "disable_tqdm": False,      # Show progress bar
-                    "logging_steps": 10,         # Print loss every 10 steps
+                    "logging_steps": 50,        # Print loss every 50 steps
                 },
             }
         },