| |
| """Numeric fidelity metrics for the multi-tool set (complement to slot accuracy): |
| num_recall : share of the reference answer's numbers present in the prediction |
| num_precision : share of the prediction's numbers that occur in the tool outputs or question |
| (a prediction number found nowhere in the context is unsupported) |
| all_backends : prediction uses at least one value unique to EACH backend block |
| python3 mt_numeric.py --ref compositional/test.jsonl --pred preds.jsonl |
| (the paper scores end-to-end outputs against compositional/test.jsonl, i.e. against the outputs of the |
| correct calls; --ref compositional/test_e2e.jsonl scores them against the outputs of the router's calls) |
| """ |
| import argparse, json, sys, os, statistics as st |
| sys.path.insert(0, os.path.dirname(os.path.abspath(__file__))) |
| from mt_qa import nums |
| ap = argparse.ArgumentParser(); ap.add_argument("--ref", required=True); ap.add_argument("--pred", required=True) |
| a = ap.parse_args() |
| ref = {json.loads(l)["question_hash"]: json.loads(l) for l in open(a.ref)} |
| rec, prec, allb, n = [], [], 0, 0 |
| for l in open(a.pred): |
| p = json.loads(l); r = ref.get(p["question_hash"]) |
| if not r: continue |
| n += 1; pr = p.get("prediction", "") |
| gn, pn = nums(r["answer"]), nums(pr) |
| ctx = nums(r["tool_result"] + "\n" + r["question"]) |
| if gn: rec.append(len(gn & pn) / len(gn)) |
| if pn: prec.append(len(pn & ctx) / len(pn)) |
| blocks = r["tool_result"].split("\n\n[") |
| sets = [nums(b) for b in blocks] |
| ok = True |
| for i, s in enumerate(sets): |
| u = s - set().union(*[t for j, t in enumerate(sets) if j != i]) - nums(r["question"]) |
| if u and not (pn & u): ok = False |
| allb += ok |
| print(json.dumps({"n": n, "num_recall": round(100 * st.mean(rec), 1), "num_precision": round(100 * st.mean(prec), 1), |
| "all_backends_used": round(100 * allb / n, 1)})) |
|
|
