Upload eval_baseline.py with huggingface_hub

eval_baseline.py

@@ -48,35 +48,31 @@ def parse(text: str):
         return None
 
 
-def walk_scenarios():
-    """Yield (scenario_id, turn_idx, atc_text, expected_intent, expected_params) for every ATC-to-ownship turn."""
+def walk_scenarios(only_ownship: bool = True):
+    """Yield (scenario_id, turn_idx, atc_text, expected_intent, expected_params) per turn.
+
+    Schema: top-level `turns[]` with {speaker, target, expected_transcript, expected_intent, expected_parameters}.
+    ATC-to-ownship subset matches `just baseline`'s 253 scored turns when only_ownship=True.
+    """
     for path in sorted(SCENARIO_DIR.glob("*.json")):
         try:
             scn = json.loads(path.read_text())
         except Exception:
             continue
-        scn_id = scn.get("id", path.stem)
-        # Walk turns — schema varies slightly across scenario types
-        # Most have "phases" with "atc_instructions" or top-level "turns" with "expected_*"
-        phases = scn.get("phases") or scn.get("dialog") or []
-        if not phases and "turns" in scn:
-            phases = [{"atc_instructions": scn.get("turns", [])}]
-        turn_idx = 0
-        for phase in phases:
-            for turn in phase.get("atc_instructions", []) + phase.get("turns", []):
-                text = turn.get("atc_text") or turn.get("text") or turn.get("transmission")
-                if not text:
-                    continue
-                # Filter to ATC-to-ownship (skip pilot replies and broadcasts to others)
-                addr = turn.get("addressed_to") or turn.get("speaker") or ""
-                if "ownship" not in str(addr).lower() and "all aircraft" not in str(addr).lower() and addr:
-                    continue
-                exp_intent = turn.get("expected_intent") or turn.get("intent")
-                exp_params = turn.get("expected_parameters") or turn.get("parameters") or {}
-                if not exp_intent:
-                    continue
-                yield scn_id, turn_idx, text, exp_intent, exp_params
-                turn_idx += 1
+        scn_id = scn.get("scenario_id", scn.get("id", path.stem))
+        for turn in scn.get("turns", []):
+            speaker = turn.get("speaker", "").lower()
+            target = turn.get("target", "").lower()
+            if speaker != "atc":
+                continue
+            if only_ownship and target != "ownship":
+                continue
+            text = turn.get("expected_transcript", "").strip()
+            exp_intent = turn.get("expected_intent")
+            exp_params = turn.get("expected_parameters") or {}
+            if not text or not exp_intent:
+                continue
+            yield scn_id, turn.get("turn_index", 0), text, exp_intent, exp_params
 
 
 def main():