Datasets:
File size: 3,359 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
"""End-to-end multi-tool evaluation helpers (router-predicted calls).
calls : python3 mt_e2e.py calls --ref splits/mt/test.jsonl --route qrm_mt_test.jsonl --out splits/mt/e2e_calls.jsonl
-> unique (design_id, tool, parameters) the router asked for, to execute with mt_ground.py
build : python3 mt_e2e.py build --ref splits/mt/test.jsonl --route qrm_mt_test.jsonl \
--results splits/mt/e2e_results.jsonl --out splits/mt/test_e2e.jsonl
-> copy of the reference set whose tool_result/stage2_context are rebuilt from the
PREDICTED calls' results (same block format and cap as mt_qa.py); the reference
answer is unchanged, so a wrong or missing call costs the answer its values.
"""
import argparse, hashlib, json, sys, os
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
from mt_qa import cap, TOTAL_CAP
def parse_calls(raw):
if not raw: return []
t = raw
i = min([k for k in (t.find("["), t.find("{")) if k >= 0], default=-1)
if i < 0: return []
try:
x, _ = json.JSONDecoder().raw_decode(t[i:])
except Exception:
return []
if isinstance(x, dict): x = [x]
out = []
for c in x if isinstance(x, list) else []:
if isinstance(c, dict) and (c.get("tool") or c.get("tool_name")):
out.append({"tool": c.get("tool") or c.get("tool_name"), "parameters": c.get("parameters") or {}})
return out[:4]
def key(did, c):
return hashlib.sha256(json.dumps([did, {"tool": c["tool"], "parameters": c["parameters"]}], sort_keys=True).encode()).hexdigest()[:16]
def main():
ap = argparse.ArgumentParser(); ap.add_argument("mode", choices=["calls", "build"])
ap.add_argument("--ref", required=True); ap.add_argument("--route", required=True)
ap.add_argument("--results"); ap.add_argument("--out", required=True)
a = ap.parse_args()
ref = [json.loads(l) for l in open(a.ref)]
route = {json.loads(l)["question_hash"]: json.loads(l)["raw"] for l in open(a.route)}
if a.mode == "calls":
seen = {}
for r in ref:
for c in parse_calls(route.get(r["question_hash"])):
k = key(r["design_id"], c); seen[k] = {"key": k, "design_id": r["design_id"], **c}
with open(a.out, "w") as f:
for c in seen.values(): f.write(json.dumps(c) + "\n")
print("unique predicted calls", len(seen)); return
res = {json.loads(l)["key"]: json.loads(l)["result"] for l in open(a.results)}
n_empty = 0
with open(a.out, "w") as f:
for r in ref:
cl = parse_calls(route.get(r["question_hash"]))
if not cl:
tr = "(no backend was called)"; n_empty += 1
else:
per = TOTAL_CAP // len(cl) - 40
tr = "\n\n".join(f"[{c['tool']} {json.dumps(c['parameters'])}]\n" +
cap(json.dumps(res.get(key(r['design_id'], c), {"status": "error", "error": "not executed"}), indent=2), per)
for c in cl)
f.write(json.dumps(dict(r, tool_result=tr, stage2_context=f"Question: {r['question']}\n\n{tr}",
predicted_calls=cl)) + "\n")
print("built", len(ref), "records;", n_empty, "with no parsable call")
if __name__ == "__main__":
main()
|