"""Bridge the public dataset and Tabulus adapters without changing Tabulus source.""" import argparse from collections import defaultdict import csv import hashlib import json from pathlib import Path, PurePosixPath DATASET="SciKnowOrg/Scientific-Table-Reconstruction" def write_json(path,value): path.parent.mkdir(parents=True,exist_ok=True) path.write_text(json.dumps(value,ensure_ascii=False,indent=2)+"\n",encoding="utf-8") def canonical_hash(value): text=json.dumps(value,sort_keys=True,ensure_ascii=False,allow_nan=False,separators=(",",":")) return hashlib.sha256(text.encode("utf-8")).hexdigest() def stage(args): directory=Path(args.dataset) if not directory.is_dir(): from huggingface_hub import snapshot_download directory=Path(snapshot_download(args.dataset,repo_type="dataset",revision="v1.0.1", allow_patterns=["manifest.json","data/*.parquet"])) import pyarrow.parquet as pq manifest=json.loads((directory/"manifest.json").read_text(encoding="utf-8")) rows=[] for file in sorted((directory/"data").glob("*.parquet")):rows.extend(pq.read_table(file).to_pylist()) rows.sort(key=lambda row:(row["paper_id"],row["table_id"])) content=[{"domain":row["domain"],"paper_id":row["paper_id"],"table_id":row["table_id"], "image_sha256":hashlib.sha256(row["image"]["bytes"]).hexdigest(), "gold_sha256":canonical_hash(row["gold_rows"])} for row in rows] if canonical_hash(content)!=manifest["dataset_fingerprint"]:raise ValueError("Dataset fingerprint mismatch") if args.out.exists() and any(args.out.iterdir()):raise ValueError("--out must be empty or new") grouped=defaultdict(list) for row in rows:grouped[row["paper_id"]].append(row) mapping=[] for pid,tables in grouped.items(): records=[] for index,row in enumerate(tables,1): relative="images/"+row["table_id"]+".png" path=args.out/pid/relative path.parent.mkdir(parents=True,exist_ok=True) data=row["image"]["bytes"] path.write_bytes(data) records.append({"table_id":index,"image":relative,"paper_id":pid,"benchmark_table_key":row["table_id"]}) mapping.append({"paper_id":pid,"table_id":row["table_id"],"tabulus_table_id":index, "image":pid+"/"+relative,"image_sha256":hashlib.sha256(data).hexdigest()}) write_json(args.out/pid/"tables_index.json",{"tables":records}) write_json(args.out/"submission_inputs.json",{"dataset":DATASET,"dataset_version":manifest["version"], "dataset_fingerprint":manifest["dataset_fingerprint"],"tables":mapping}) print(f"Staged {len(rows)} original images from {len(grouped)} papers; no gold rows were written.") def collect(args): mapping=json.loads((args.inputs/"submission_inputs.json").read_text(encoding="utf-8")) expected=defaultdict(dict) for table in mapping["tables"]: data=(args.inputs/table["image"]).read_bytes() if hashlib.sha256(data).hexdigest()!=table["image_sha256"]:raise ValueError("An inference input changed") expected[table["paper_id"]][table["tabulus_table_id"]]=table["table_id"] output=[];audit=[] saved=getattr(args,"saved_dataset",None) saved_papers={} if saved: with (saved/"metadata"/"tables.csv").open(encoding="utf-8-sig",newline="") as handle: for row in csv.DictReader(handle): parts=Path(row["gold_table_path"]).parts saved_papers[row["paper_id"]]=saved.joinpath(*parts[:parts.index("step2_table_reconstruction")]) for pid,indices in sorted(expected.items()): directory=(saved_papers[pid]/"tabulus_runs"/"step2"/"controlled"/args.run/args.adapter/"reconstruction" if saved else args.outputs/pid/args.adapter) manifest_path=directory/"batch_summary.json" if not manifest_path.exists():raise ValueError(f"Missing run manifest for {pid}; do not count unattempted tables as model errors") summary=json.loads(manifest_path.read_text(encoding="utf-8")) if summary["adapter_name"]!=args.adapter:raise ValueError(f"Wrong adapter for {pid}") items=summary["items"] if saved: observed={} for item in items: source=PurePosixPath(item["source_image"].replace("\\","/")) key=source.parent.name if source.stem=="crop" else source.stem if key in observed:raise ValueError(f"Duplicate source identity for {pid}/{key}") observed[key]=item if set(observed)!=set(indices.values()):raise ValueError(f"Saved source IDs differ from released inputs for {pid}") else: observed={item["table_id"]:item for item in items} if len(observed)!=len(items) or set(observed)!=set(indices):raise ValueError(f"Run IDs differ from staged inputs for {pid}") for number,key in indices.items(): item=observed[key] if saved else observed[number] status=item["status"] rows=[];note=None if status not in ("ok","empty","error"):raise ValueError(f"Unexpected status: {status}") if status=="ok" and item.get("prediction_csv"): path=(directory/item["prediction_csv"]).resolve() if not path.is_relative_to(directory.resolve()):raise ValueError("Prediction path escapes its run folder") if not path.is_file():raise ValueError(f"Missing saved prediction CSV for {pid}/{key}") with path.open(encoding="utf-8-sig",newline="") as handle:rows=list(csv.reader(handle)) if not any(cell.strip() for row in rows for cell in row):status="empty";rows=[] elif status=="ok": status="empty";note="Adapter completed but produced no table CSV" output.append({"paper_id":pid,"table_id":key,"status":status,"rows":rows}) audit.append({"paper_id":pid,"table_id":key,"adapter_status":item["status"], "submission_status":status,"note":note}) args.out.parent.mkdir(parents=True,exist_ok=True) if args.out.exists():raise ValueError("Refusing to overwrite an existing predictions file") output.sort(key=lambda row:(row["paper_id"],row["table_id"])) args.out.write_text("".join(json.dumps(row,ensure_ascii=False)+"\n" for row in output),encoding="utf-8") write_json(args.out.with_suffix(".audit.json"),{"dataset_fingerprint":mapping["dataset_fingerprint"], "adapter":args.adapter,"tables":audit}) from collections import Counter print(json.dumps({"predictions":len(output),"status":Counter(row["status"] for row in output)})) def main(): parser=argparse.ArgumentParser(description=__doc__) commands=parser.add_subparsers(dest="command",required=True) prepare=commands.add_parser("stage",help="Write image-only canonical crop folders accepted by Tabulus") prepare.add_argument("--dataset",default=DATASET,help="HF dataset ID or local release directory") prepare.add_argument("--out",type=Path,required=True) prepare.set_defaults(function=stage) export=commands.add_parser("collect",help="Package the Tabulus batch outputs as leaderboard predictions") export.add_argument("--inputs",type=Path,required=True) source=export.add_mutually_exclusive_group(required=True) source.add_argument("--outputs",type=Path,help="Output parent from a new Tabulus --crops-folder run") source.add_argument("--saved-dataset",type=Path,help="Existing TabulusBench tree containing controlled saved runs") export.add_argument("--run",default="run_01",help="Explicit saved controlled run to reuse with --saved-dataset") export.add_argument("--adapter",default="chandra") export.add_argument("--out",type=Path,required=True) export.set_defaults(function=collect) args=parser.parse_args() args.function(args) if __name__=="__main__":main()