"""Next Trace inference (NumPy only): evidence sentences or abstention for a question over a document. Usage: python inference.py --selftest python inference.py --doc "..." --question "..." """ import argparse import json import os import sys HERE = os.path.dirname(os.path.abspath(__file__)) sys.path.insert(0, HERE) import trace_core as tc # noqa: E402 from nlrt import load_model # noqa: E402 def load(path=os.path.join(HERE, "model.json")): return load_model(path) def trace(model, question, document): r = tc.score_document(model, question, document) return {"abstain": r["abstain"], "answerable": r["answerable"], "evidence": [{"index": i, "sentence": document[r["spans"][i][0]:r["spans"][i][1]], "p": r["probs"][i]} for i in r["evidence"]]} def _selftest(): model = load() with open(os.path.join(HERE, "selftest.json"), encoding="utf-8") as f: cases = json.load(f)["cases"] worst, bad = 0.0, [] for c in cases: r = tc.score_document(model, c["question"], c["document"]) worst = max([worst] + [abs(a - b) for a, b in zip(r["probs"], c["probs"])] + [abs(r["answerable"] - c["answerable"])]) if r["abstain"] != c["abstain"] or r["evidence"] != c["evidence"]: bad.append(c["question"]) ok = not bad and worst < 1e-6 print(json.dumps({"passed": ok, "cases": len(cases), "max_abs_diff": worst, "decision_mismatches": bad}, ensure_ascii=False)) return ok if __name__ == "__main__": ap = argparse.ArgumentParser() ap.add_argument("--selftest", action="store_true") ap.add_argument("--doc", default="") ap.add_argument("--question", default="") a = ap.parse_args() if a.selftest: sys.exit(0 if _selftest() else 1) print(json.dumps(trace(load(), a.question, a.doc), ensure_ascii=False, indent=1))