scripts/build_dataset.py

#!/usr/bin/env python3
"""Build the public UrbanGround Hugging Face dataset from App task JSON files."""

from __future__ import annotations

import argparse
import copy
import json
import re
from collections import Counter
from pathlib import Path
from typing import Any


TYPE_INFO = {
    0: ("Short-Range Goal Navigation", "SGN", 2, "Navigation under Explicit Instructions"),
    2: ("Long-Range Goal Navigation", "LGN", 2, "Navigation under Explicit Instructions"),
    3: ("Place-Type Search", "PTS", 3, "Exploration under Implicit Instructions"),
    5: ("Time-Window Scheduling", "TWS", 4, "Multi-Task Planning"),
    7: ("Visual Recognition", "VR", 1, "Local Environment Understanding"),
    8: ("Orientation Understanding", "OU", 1, "Local Environment Understanding"),
    9: ("Active Exploration Questions", "AEQ", 1, "Local Environment Understanding"),
    10: ("Instructional Navigation", "IN", 2, "Navigation under Explicit Instructions"),
    11: ("Constrained Navigation", "CN", 2, "Navigation under Explicit Instructions"),
    12: ("Implicit Intent Inference", "III", 3, "Exploration under Implicit Instructions"),
    13: ("Multi-Stop Route Planning", "MSP", 4, "Multi-Task Planning"),
    14: ("Dynamic Road-Closure Replanning", "DCR", 5, "Dynamic Environment Interaction"),
    15: ("Navigation among Pedestrians", "NP", 5, "Dynamic Environment Interaction"),
}

BASE_PREFIX_TO_TYPE = {
    "LQ": 7,
    "OQ": 8,
    "SQ": 9,
    "SN": 0,
    "LN": 2,
    "IN": 10,
    "CN": 11,
    "PS": 3,
    "II": 12,
    "SF": 5,
    "MP": 13,
}

POI_CATEGORIES = {
    0: "Education",
    1: "Medical & Health",
    2: "Social Welfare",
    3: "Recreation & Sports",
    4: "Open Space",
    5: "Culture & Entertainment",
    6: "Religious & Burial Facilities",
    7: "Municipal & Public Utilities",
    8: "Government Offices",
    9: "Transport",
    10: "Commercial & Retail",
    11: "Food & Beverage",
    12: "Accommodation",
    13: "Residential",
}

TYPE_ORDER = [7, 8, 9, 0, 2, 10, 11, 3, 12, 5, 13, 14, 15]
TYPE_ORDER_INDEX = {task_type: index for index, task_type in enumerate(TYPE_ORDER)}

EXPECTED_BASE_COUNTS = {
    7: 80,
    8: 60,
    9: 80,
    0: 80,
    2: 80,
    10: 50,
    11: 30,
    3: 60,
    12: 60,
    5: 60,
    13: 60,
}
EXPECTED_BENCHMARK_COUNTS = {
    **EXPECTED_BASE_COUNTS,
    14: 30,
    15: 80,
}


def snake_case(name: str) -> str:
    first = re.sub(r"(.)([A-Z][a-z]+)", r"\1_\2", name)
    return re.sub(r"([a-z0-9])([A-Z])", r"\1_\2", first).lower()


def snake_case_keys(value: Any) -> Any:
    if isinstance(value, dict):
        return {snake_case(str(key)): snake_case_keys(item) for key, item in value.items()}
    if isinstance(value, list):
        return [snake_case_keys(item) for item in value]
    return value


def read_task_file(path: Path) -> dict[str, Any]:
    with path.open("r", encoding="utf-8") as handle:
        document = json.load(handle)
    if set(document) != {"task"} or not isinstance(document["task"], dict):
        raise ValueError(f"{path}: expected one top-level 'task' object")
    task = document["task"]
    if task.get("id") != path.stem:
        raise ValueError(f"{path}: filename does not match task id {task.get('id')!r}")
    if not isinstance(task.get("type"), int):
        raise ValueError(f"{path}: task type must be an integer")
    return task


def make_record(task: dict[str, Any]) -> dict[str, Any]:
    task_type_id = int(task["type"])
    if task_type_id not in TYPE_INFO:
        raise ValueError(f"{task['id']}: unsupported task type {task_type_id}")

    task_type, abbreviation, level, capability = TYPE_INFO[task_type_id]
    payload = snake_case_keys(copy.deepcopy(task))
    task_id = str(payload.pop("id"))
    payload.pop("type")
    payload.pop("source_task_id", None)

    qa_answer_text = ""
    qa_options = payload.get("qa_options", [])
    qa_answer_index = payload.get("qa_answer_index")
    if isinstance(qa_answer_index, int) and 0 <= qa_answer_index < len(qa_options):
        qa_answer_text = str(qa_options[qa_answer_index].get("text", ""))

    poi_category = payload.get("poi_category")
    poi_category_name = (
        POI_CATEGORIES.get(poi_category, "") if task_type_id == 3 else ""
    )

    record: dict[str, Any] = {
        "id": task_id,
        "task_type_id": task_type_id,
        "task_type": task_type,
        "task_abbreviation": abbreviation,
        "capability_level": level,
        "capability_name": capability,
        "qa_answer_text": qa_answer_text,
        "poi_category_name": poi_category_name,
    }
    record.update(payload)
    return record


def make_level_five_task(source: dict[str, Any], target_type: int) -> dict[str, Any]:
    if target_type == 14 and source["type"] != 11:
        raise ValueError("DCR requires constrained-navigation task geometry")
    if target_type == 15 and source["type"] != 2:
        raise ValueError("NP requires long-range navigation task geometry")

    task = copy.deepcopy(source)
    abbreviation = TYPE_INFO[target_type][1]
    _, separator, suffix = str(source["id"]).partition("-")
    if not separator:
        raise ValueError(f"{source['id']}: expected a prefixed task id")
    task["id"] = f"{abbreviation}-{suffix}"
    task["type"] = target_type
    task["sourceTaskId"] = ""
    return task


def write_jsonl(path: Path, records: list[dict[str, Any]]) -> None:
    with path.open("w", encoding="utf-8", newline="\n") as handle:
        for record in records:
            handle.write(json.dumps(record, ensure_ascii=False, separators=(",", ":")))
            handle.write("\n")


def build(source_dir: Path, output_dir: Path) -> None:
    task_paths = sorted(source_dir.glob("*.json"), key=lambda path: path.name)
    if not task_paths:
        raise ValueError(f"no task JSON files found in {source_dir}")

    stored_tasks = [(path, read_task_file(path)) for path in task_paths]
    ids = [task["id"] for _, task in stored_tasks]
    if len(ids) != len(set(ids)):
        raise ValueError("duplicate task ids found")

    benchmark_source_tasks: list[dict[str, Any]] = []
    for path, task in stored_tasks:
        prefix = task["id"].split("-", 1)[0]
        expected_type = BASE_PREFIX_TO_TYPE.get(prefix)
        if expected_type is None:
            continue
        if task["type"] != expected_type:
            raise ValueError(
                f"{task['id']}: prefix expects type {expected_type}, got {task['type']}"
            )
        benchmark_source_tasks.append(task)

    base_counts = Counter(task["type"] for task in benchmark_source_tasks)
    if dict(base_counts) != EXPECTED_BASE_COUNTS:
        raise ValueError(
            f"base task counts changed: expected {EXPECTED_BASE_COUNTS}, got {dict(base_counts)}"
        )

    benchmark_records: list[dict[str, Any]] = []
    for task in benchmark_source_tasks:
        benchmark_records.append(make_record(task))

    for task in benchmark_source_tasks:
        if task["type"] == 11:
            benchmark_records.append(make_record(make_level_five_task(task, 14)))
        elif task["type"] == 2:
            benchmark_records.append(make_record(make_level_five_task(task, 15)))

    benchmark_records.sort(
        key=lambda record: (
            record["capability_level"],
            TYPE_ORDER_INDEX[record["task_type_id"]],
            record["id"],
        )
    )
    for index, record in enumerate(benchmark_records):
        record["instance_index"] = index

    benchmark_type_counts = Counter(record["task_type_id"] for record in benchmark_records)
    if dict(benchmark_type_counts) != EXPECTED_BENCHMARK_COUNTS:
        raise ValueError(
            "benchmark task counts changed: "
            f"expected {EXPECTED_BENCHMARK_COUNTS}, got {dict(benchmark_type_counts)}"
        )
    if len(benchmark_records) != 810:
        raise ValueError(f"expected 810 benchmark instances, got {len(benchmark_records)}")

    data_dir = output_dir / "data"
    metadata_dir = output_dir / "metadata"
    data_dir.mkdir(parents=True, exist_ok=True)
    metadata_dir.mkdir(parents=True, exist_ok=True)

    write_jsonl(data_dir / "benchmark.jsonl", benchmark_records)

    benchmark_counts = Counter(
        (record["capability_level"], record["task_abbreviation"])
        for record in benchmark_records
    )
    statistics = {
        "total_instances": len(benchmark_records),
        "splits": {"test": len(benchmark_records)},
        "benchmark_counts": [
            {
                "capability_level": level,
                "task_abbreviation": abbreviation,
                "number": count,
            }
            for (level, abbreviation), count in sorted(benchmark_counts.items())
        ],
    }
    (metadata_dir / "statistics.json").write_text(
        json.dumps(statistics, ensure_ascii=False, indent=2, sort_keys=True) + "\n",
        encoding="utf-8",
    )

    print(f"Built {len(benchmark_records)} UrbanGround task records")


def parse_args() -> argparse.Namespace:
    script_dir = Path(__file__).resolve().parent
    default_output = script_dir.parent
    default_source = default_output.parents[1] / "task"
    parser = argparse.ArgumentParser()
    parser.add_argument("--source", type=Path, default=default_source)
    parser.add_argument("--output", type=Path, default=default_output)
    return parser.parse_args()


if __name__ == "__main__":
    arguments = parse_args()
    build(arguments.source.resolve(), arguments.output.resolve())