results: select BabyAI clean configuration

1 files changed, 86 insertions, 0 deletions

diff --git a/experiments/analyze_babyai_shared_b0.py b/experiments/analyze_babyai_shared_b0.py
new file mode 100644
index 0000000..98cfae8
--- /dev/null
+++ b/experiments/analyze_babyai_shared_b0.py
@@ -0,0 +1,86 @@
+#!/usr/bin/env python3
+"""Apply the frozen BabyAI B0 clean-selector rule."""
+
+import argparse
+import json
+from pathlib import Path
+
+
+ROOT = Path(__file__).resolve().parents[1]
+DEFAULT_RESULTS = ROOT / "results" / "babyai_shared" / "b0"
+DEFAULT_OUT = ROOT / "results" / "babyai_shared" / "b0_selector.json"
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--results", type=Path, default=DEFAULT_RESULTS)
+    parser.add_argument("--out", type=Path, default=DEFAULT_OUT)
+    args = parser.parse_args()
+    rows = []
+    for depth in (2, 4):
+        for learning_rate in (0.01, 0.03):
+            records = {}
+            for condition in ("bp", "clean_kp"):
+                path = args.results / (
+                    f"d{depth}_lr{learning_rate}_{condition}.json")
+                with open(path, encoding="utf-8") as handle:
+                    records[condition] = json.load(handle)
+            bp = records["bp"]
+            kp = records["clean_kp"]
+            bp_success = float(bp["rollout"]["success"])
+            kp_success = float(kp["rollout"]["success"])
+            bp_lesion = float(bp["mission_lesion_rollout"]["success"])
+            checks = {
+                "both_finite": bool(bp["finite"] and kp["finite"]),
+                "bp_success_at_least_0p8": bp_success >= 0.8,
+                "clean_kp_success_at_least_0p8": kp_success >= 0.8,
+                "bp_mission_lesion_drop_at_least_0p2": (
+                    bp_success - bp_lesion >= 0.2),
+            }
+            rows.append({
+                "hidden_layers": depth,
+                "learning_rate": learning_rate,
+                "bp_rollout_success": bp_success,
+                "clean_kp_rollout_success": kp_success,
+                "bp_mission_lesion_success": bp_lesion,
+                "bp_mission_lesion_drop": bp_success - bp_lesion,
+                "bp_action_accuracy": float(bp["validation"]["accuracy"]),
+                "clean_kp_action_accuracy": float(
+                    kp["validation"]["accuracy"]),
+                "eligible": all(checks.values()),
+                "checks": checks,
+                "source_files": [str(
+                    (args.results / f"d{depth}_lr{learning_rate}_{condition}.json")
+                    .relative_to(ROOT)) for condition in ("bp", "clean_kp")],
+            })
+    eligible = [row for row in rows if row["eligible"]]
+    selected = max(eligible, key=lambda row: (
+        row["clean_kp_rollout_success"], row["clean_kp_action_accuracy"],
+        -row["hidden_layers"], -row["learning_rate"])) if eligible else None
+    report = {
+        "stage": "babyai_shared_b0_selector",
+        "gate": "pass" if selected is not None else "fail",
+        "selection_rule": (
+            "highest clean-KP rollout success, then action accuracy, then "
+            "fewer layers, then smaller learning rate, among eligible rows"),
+        "candidates": rows,
+        "selected": ({
+            "hidden_layers": selected["hidden_layers"],
+            "width": 256,
+            "learning_rate": selected["learning_rate"],
+            "context_gain": 1.0,
+            "b1_epochs": 40,
+            "b1_model_and_shuffle_seeds": [4101, 4102, 4103],
+        } if selected is not None else None),
+        "raw_or_sdil_results_read": False,
+        "test_split_generated_or_read": False,
+    }
+    args.out.parent.mkdir(parents=True, exist_ok=True)
+    with open(args.out, "w", encoding="utf-8") as handle:
+        json.dump(report, handle, indent=2, sort_keys=True)
+        handle.write("\n")
+    print(json.dumps(report, indent=2, sort_keys=True))
+
+
+if __name__ == "__main__":
+    main()