Datasets:
File size: 3,665 Bytes
676defa | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 | #!/usr/bin/env python3
"""Score the multi-tool set with the SAME slot and recall functions as score.py.
python3 mt_score.py --ref compositional/test.jsonl --pred preds.jsonl [--route qrm_preds.jsonl] [--out-json x.json]
Answer metrics (per 2-/3-backend group and overall): slot accuracy = every number
in the reference answer and the module name appear in the prediction; rec80;
mean content recall. Routing (optional): the predicted call list must equal the
reference call set (tool and parameters, order ignored); also reports
backend-set accuracy ignoring parameters.
"""
import argparse, json, collections, sys, os
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
from score import slot_match, recall, wilson
def calls_of(x):
if isinstance(x, str):
from mt_e2e import parse_calls
x = parse_calls(x) or None
if isinstance(x, dict): x = [x]
if not isinstance(x, list): return None
try: return sorted(json.dumps({"tool": c["tool"], "parameters": c.get("parameters") or {}}, sort_keys=True) for c in x)
except Exception: return None
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--ref", required=True); ap.add_argument("--pred"); ap.add_argument("--route")
ap.add_argument("--out-json")
a = ap.parse_args()
ref = {json.loads(l)["question_hash"]: json.loads(l) for l in open(a.ref)}
out = {"ref": a.ref, "n": len(ref)}
if a.pred:
preds = {}
for l in open(a.pred):
d = json.loads(l); preds[d["question_hash"]] = d.get("prediction", "")
g = collections.defaultdict(lambda: collections.Counter())
for h, r in ref.items():
for key in ("all", f"{r['n_calls']}-backend"):
p = preds.get(h, ""); s = g[key]; s["n"] += 1
sm = slot_match(p, r["answer"], r["module_name"])
if sm is not None: s["slot_n"] += 1; s["slot_ok"] += bool(sm)
rc = recall(p, r["answer"])
if rc is not None: s["rec_n"] += 1; s["rec_sum"] += rc; s["rec80"] += rc >= 0.8
out["missing"] = len(set(ref) - set(preds)); out["answer"] = {}
for k, s in sorted(g.items()):
lo, hi = wilson(s["slot_ok"], s["slot_n"])
out["answer"][k] = {"n": s["n"], "slot_ok": s["slot_ok"], "slot_n": s["slot_n"],
"slot": 100 * s["slot_ok"] / max(1, s["slot_n"]), "slot_ci": [lo, hi],
"rec80": 100 * s["rec80"] / max(1, s["rec_n"]), "recall": 100 * s["rec_sum"] / max(1, s["rec_n"])}
print(f"{k:10} n={s['n']:4} slot {s['slot_ok']}/{s['slot_n']} = {out['answer'][k]['slot']:.1f}% [{lo:.1f}, {hi:.1f}]"
f" rec80 {out['answer'][k]['rec80']:.1f}% recall {out['answer'][k]['recall']:.1f}%")
if a.route:
rp = {}
for l in open(a.route):
d = json.loads(l); rp[d["question_hash"]] = d.get("raw", d.get("calls"))
exact = tools_only = 0
for h, r in ref.items():
gt = calls_of(r["stage1_target"]); pr = calls_of(rp.get(h))
exact += pr == gt
tools_only += pr is not None and sorted(json.loads(c)["tool"] for c in pr) == sorted(json.loads(c)["tool"] for c in gt)
n = len(ref); lo, hi = wilson(exact, n)
out["routing"] = {"exact": exact, "tools_only": tools_only, "n": n, "exact_pct": 100 * exact / n, "ci": [lo, hi]}
print(f"routing exact {exact}/{n} = {100*exact/n:.1f}% [{lo:.1f}, {hi:.1f}] backend set {tools_only}/{n} = {100*tools_only/n:.1f}%")
if a.out_json: json.dump(out, open(a.out_json, "w"), indent=1)
if __name__ == "__main__":
main()
|