| |
| """Honest input-only baselines. The symbolic baseline recognizes authored templates.""" |
| import argparse |
| import collections |
| import json |
| import pathlib |
| from model_input import model_input |
| from verify import evaluate, read_lines |
|
|
|
|
| def symbolic(item): |
| question, cells = item["question_ar"], item["table"]["cells"] |
| if question.startswith("هل يمكن جمع"): |
| op = "incompatible_sum" |
| elif "اقسم إجمالي الناتج المحلي" in question: |
| op = "gdp_per_capita_join" |
| elif question.startswith("بعد ترتيب القيم المختلفة"): |
| op = "second_highest_years" |
| elif "القيمة الأعلى" in question: |
| op = "maximum_years" |
| elif "القيمة الأدنى" in question: |
| op = "minimum_years" |
| elif "نسبة التغير" in question: |
| op = "relative_change" |
| elif "التغير بالنقاط المئوية" in question: |
| op = "percentage_point_difference" |
| elif "الفرق بالوحدة المبينة" in question: |
| op = "difference" |
| elif "أكبر من" in question: |
| op = "greater_than" |
| elif "ما القيمة المسجلة" in question: |
| op = "lookup" |
| else: |
| raise ValueError("Unsupported authored template") |
| return evaluate(op, cells) |
|
|
|
|
| def first_cell(item): |
| row = item["table"]["cells"][0] |
| return {"status": "missing" if row["value"] is None else "ok", "value": row["value"], "unit": row["unit_code"]} |
|
|
|
|
| def run(release): |
| questions = sum([read_lines(release / "jsonl" / "questions" / (split + ".jsonl")) for split in ("train", "validation", "test")], []) |
| report = {"learned_model_evaluation": False, |
| "scope": "Full released rows, using only model_input(question). The symbolic method recognizes every declared authored Arabic template and is a programmed format ceiling, not general Arabic understanding, training benefit or model performance.", "baselines": {}} |
| for name, method in [("first_visible_cell", first_cell), ("template_aware_symbolic", symbolic)]: |
| stats = collections.defaultdict(lambda: {"rows": 0, "exact_answer": 0, "status_exact": 0}) |
| predictions = [] |
| for q in questions: |
| prediction = method(model_input(q)) |
| gold = json.loads(q["gold_answer"]) |
| for key in ("all", "split:" + q["split"], "operation:" + q["operation"], "diagnostic" if q["structure_holdout_diagnostic"] else "shared_operators"): |
| stats[key]["rows"] += 1 |
| stats[key]["exact_answer"] += prediction == gold |
| stats[key]["status_exact"] += prediction["status"] == gold["status"] |
| evidence = [row["cell_id"] for row in model_input(q)["table"]["cells"]] |
| predictions.append({"id": q["id"], "answer": prediction, |
| "evidence_ids": evidence if name == "template_aware_symbolic" else evidence[:1]}) |
| for item in stats.values(): |
| item["exact_answer_rate"] = item["exact_answer"] / item["rows"] |
| item["status_exact_rate"] = item["status_exact"] / item["rows"] |
| report["baselines"][name] = dict(stats) |
| dest = release / "reports" / (name + ".predictions.jsonl") |
| dest.parent.mkdir(parents=True, exist_ok=True) |
| dest.write_text("".join(json.dumps(x, ensure_ascii=False, sort_keys=True) + "\n" for x in predictions)) |
| return report |
|
|
|
|
| if __name__ == "__main__": |
| parser = argparse.ArgumentParser() |
| parser.add_argument("--release", type=pathlib.Path, required=True) |
| args = parser.parse_args() |
| report = run(args.release) |
| (args.release / "reports" / "baselines.json").write_text(json.dumps(report, indent=2) + "\n") |
| print(json.dumps(report, indent=2)) |
|
|
