| |
| """Plants one deliberate mistake per scenario family in a small copy of the bench; the validator must flag every one. |
| usage: mutation_test.py <bench dir> <original airline db.json> <out dir>""" |
| import copy, json, random, sys |
| bench, orig_db_path, out = sys.argv[1:4] |
| tasks = json.load(open(bench + "/tasks.json")); db = json.load(open(bench + "/db.json")); orig = json.load(open(orig_db_path)) |
| fam = lambda t: json.loads(t["description"]["notes"])["family"] |
| first = {} |
| for t in tasks: |
| first.setdefault(fam(t), t) |
| rng = random.Random(1) |
| def acts(t): return t["evaluation_criteria"]["actions"] |
| def m_cancel_denied(t): acts(t).append({"action_id": "x", "name": "cancel_reservation", "arguments": {"reservation_id": json.loads(t["description"]["notes"])["reservation"]}, "info": None}) |
| def m_cancel_ok(t): acts(t).clear() |
| def m_baggage(t): acts(t)[0]["arguments"]["nonfree_baggages"] += 1 |
| def m_booking(t): acts(t)[0]["arguments"]["payment_methods"][0]["amount"] -= 10 |
| def m_flight(t): |
| a = acts(t)[0]["arguments"]; cur = db["flights"][a["flights"][0]["flight_number"]] |
| other = next(f for f in db["flights"].values() if f["origin"] != cur["origin"] and a["flights"][0]["date"] in f["dates"] and f["dates"][a["flights"][0]["date"]]["status"] == "available") |
| a["flights"][0]["flight_number"] = other["flight_number"] |
| def m_sweep(t): acts(t).pop() if acts(t) else acts(t).append({"action_id": "x", "name": "cancel_reservation", "arguments": {"reservation_id": db["users"][json.loads(t["description"]["notes"])["users"][0]]["reservations"][0]}, "info": None}) |
| def m_comp(t): acts(t)[0]["arguments"]["amount"] = 50 |
| def m_pax(t): acts(t)[0]["arguments"]["passengers"].append({"first_name": "Extra", "last_name": "Person", "dob": "1990-01-01"}) |
| def m_cabin(t): acts(t)[0]["arguments"]["cabin"] = db["reservations"][acts(t)[0]["arguments"]["reservation_id"]]["cabin"] |
| def m_multi(t): acts(t)[1]["arguments"]["passengers"].pop() if len(acts(t)[1]["arguments"]["passengers"]) > 1 else acts(t)[1]["arguments"]["passengers"].append({"first_name": "Extra", "last_name": "Person", "dob": "1990-01-01"}) |
| def m_leak(t): t["user_scenario"]["instructions"]["known_info"] += "\n\nYour old user id was %s." % sorted(orig["users"])[0] |
| def m_rule(t): |
| a = acts(t)[0]["arguments"]; day = a["flights"][0]["date"] |
| alt = next(f for f in db["flights"].values() if (f["origin"], f["destination"]) == (a["origin"], a["destination"]) and f["flight_number"] != a["flights"][0]["flight_number"] and f["dates"].get(day, {}).get("status") == "available") |
| old = db["flights"][a["flights"][0]["flight_number"]]["dates"][day]["prices"][a["cabin"]]; new = alt["dates"][day]["prices"][a["cabin"]] |
| a["flights"][0]["flight_number"] = alt["flight_number"] |
| a["payment_methods"][-1]["amount"] += (new - old) * len(a["passengers"]) |
| plan = {"cancel_denied": m_cancel_denied, "cancel_ok": m_cancel_ok, "baggage_add": m_baggage, "booking": m_booking, "flight_change": m_flight, |
| "sweep_cancel": m_sweep, "sweep_cabin": m_sweep, "compensation_ok": m_comp, "passenger_update": m_pax, "cabin_change": m_cabin, |
| "multi_change": m_multi, "compensation_false_claim": m_leak, "booking_by_rule": m_rule} |
| bad, clean = [], [] |
| for f, fn in plan.items(): |
| t = copy.deepcopy(first[f]); fn(t); bad.append(t) |
| for f, t in first.items(): |
| if f not in plan: clean.append(copy.deepcopy(t)) |
| all_t = bad + clean |
| json.dump(all_t, open(out + "/tasks.json", "w")); json.dump(db, open(out + "/db.json", "w")) |
| ids = [t["id"] for t in all_t] |
| json.dump({"base": ids, "train": ids[::2], "holdout": ids[1::2], "devil": ids[1::2]}, open(out + "/split_tasks.json", "w")) |
| json.dump([t["id"] for t in bad], open(out + "/planted.json", "w")) |
| print("planted mistakes in:", ", ".join(t["id"] for t in bad)) |
|
|
