load_dataset.py

"""
FCC Invoices Verified Augmented - Dataset Loader

Usage:
    from huggingface_hub import snapshot_download
    from load_dataset import FCCInvoicesDataset

    local_dir = snapshot_download(repo_id="amazon/ConfBench", repo_type="dataset")
    ds = FCCInvoicesDataset(local_dir=local_dir)

    # Iterate all (document, pipeline) pairs
    for sample in ds:
        print(sample["doc_id"], sample["pipeline_name"])
        gt = sample["ground_truth"]["inference_result"]
        print(gt["Agency"], gt["GrossTotal"])

    # Get the local path to a single noisy PDF
    path = ds.pdf_path(doc_id="033f718b16cb597c065930410752c294", pipeline_name="custom13")

Requirements:
    pip install huggingface_hub
"""

import json
from pathlib import Path
from typing import Iterator


PIPELINES = [
    "default", "archetype3", "archetype4", "archetype7", "archetype9",
    "archetype10", "archetype11", "custom12", "custom13", "custom14",
    "custom15", "custom16", "custom17", "custom18", "custom19",
    "custom20", "custom21", "custom22",
]


class FCCInvoicesDataset:
    """Lazy iterator over all (document, pipeline) pairs in a local ConfBench checkout."""

    def __init__(self, local_dir: str):
        self.local_dir = Path(local_dir)
        # Documents live under assets/; accept a checkout root or the assets dir itself.
        if (self.local_dir / "assets").is_dir():
            self.assets_dir = self.local_dir / "assets"
        else:
            self.assets_dir = self.local_dir
        self._doc_ids: list[str] | None = None

    # ------------------------------------------------------------------
    # Public API
    # ------------------------------------------------------------------

    def doc_ids(self) -> list[str]:
        """Return all document IDs (md5 hashes)."""
        if self._doc_ids is None:
            self._doc_ids = sorted(
                p.name for p in self.assets_dir.iterdir()
                if p.is_dir() and (p / "metadata.json").exists()
            )
        return self._doc_ids

    def load_ground_truth(self, doc_id: str) -> dict:
        """Read and parse gt.json for a document."""
        return json.loads((self.assets_dir / doc_id / "gt.json").read_text())

    def load_metadata(self, doc_id: str) -> dict:
        """Read and parse metadata.json (pipeline manifest) for a document."""
        return json.loads((self.assets_dir / doc_id / "metadata.json").read_text())

    def pdf_path(self, doc_id: str, pipeline_name: str) -> str:
        """Return the local path to a noisy PDF."""
        return str(self.assets_dir / doc_id / pipeline_name / f"{pipeline_name}_noisy.pdf")

    def original_pdf_path(self, doc_id: str) -> str:
        """Return the local path to the original clean PDF."""
        return str(self.assets_dir / doc_id / "original.pdf")

    def get_sample(self, doc_id: str, pipeline_name: str) -> dict:
        """Return a single sample dict."""
        gt = self.load_ground_truth(doc_id)
        return {
            "doc_id":        doc_id,
            "pipeline_name": pipeline_name,
            "pipeline_type": "archetype" if not pipeline_name.startswith("custom") else "custom",
            "original_pdf":  self.original_pdf_path(doc_id),
            "noisy_pdf":     self.pdf_path(doc_id, pipeline_name),
            "ground_truth":  gt,
        }

    def __iter__(self) -> Iterator[dict]:
        """Yield one dict per (document, pipeline) pair."""
        for doc_id in self.doc_ids():
            meta = self.load_metadata(doc_id)
            gt   = self.load_ground_truth(doc_id)
            for pipeline in meta.get("pipelines", []):
                name = pipeline["pipeline_name"]
                yield {
                    "doc_id":        doc_id,
                    "pipeline_name": name,
                    "pipeline_type": "archetype" if not name.startswith("custom") else "custom",
                    "original_pdf":  self.original_pdf_path(doc_id),
                    "noisy_pdf":     self.pdf_path(doc_id, name),
                    "ground_truth":  gt,
                }

    def __len__(self) -> int:
        return len(self.doc_ids()) * len(PIPELINES)


# ------------------------------------------------------------------
# CLI convenience
# ------------------------------------------------------------------
if __name__ == "__main__":
    import argparse

    parser = argparse.ArgumentParser(description="FCC Invoices dataset helper")
    parser.add_argument("--local-dir", default=".", help="Path to local ConfBench checkout")
    sub = parser.add_subparsers(dest="cmd")

    p_list = sub.add_parser("list", help="List all document IDs")

    p_gt = sub.add_parser("gt", help="Print ground truth for a document")
    p_gt.add_argument("doc_id")

    p_path = sub.add_parser("path", help="Print the local path to a noisy PDF")
    p_path.add_argument("doc_id")
    p_path.add_argument("pipeline_name")

    args = parser.parse_args()
    ds = FCCInvoicesDataset(local_dir=args.local_dir)

    if args.cmd == "list":
        for d in ds.doc_ids():
            print(d)
    elif args.cmd == "gt":
        print(json.dumps(ds.load_ground_truth(args.doc_id), indent=2))
    elif args.cmd == "path":
        print(ds.pdf_path(args.doc_id, args.pipeline_name))
    else:
        parser.print_help()