File size: 4,041 Bytes
31caa7f | 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 69 70 71 72 73 74 75 76 | """How did runs differ by model? Source profile of every research move, mechanically classified by tool signature.
Permutation test: shuffle model labels within each (task, repeat) triple, so task and round are held fixed."""
import json, re, random, statistics as st, collections, sys
from pathlib import Path
B = Path(sys.argv[1])
H = re.compile(r"^(draft:|shell:inspect|help:|skills|catalog|file:|shell:other)")
GOV = re.compile(r"(\.gov$|\.gov\.|europa\.eu$|ecb\.|\.int$|sec\.gov)")
def kind(sig):
s = sig
if re.match(r"^(websearch|browser:search|pplx)", s): return "search"
if re.match(r"^(sec-filings|trade|funding|fda|biotech)", s): return "data"
if re.match(r"^(shell:process|script|xlsx|describe_image|workbooks)", s): return "process"
m = re.match(r"^(webfetch|pdftext|doctext|browser:(asset|session|fetch|get|show)|curl)(:(.+))?$", s)
if m:
d = (m.group(4) or "")
return "official" if GOV.search(d) else "page"
return "other"
KINDS = ["search", "page", "official", "data", "process"]
moves = collections.defaultdict(list)
for l in open(B / "moves.jsonl"):
m = json.loads(l)
sigs = [x for x in m["signature"].split(" + ") if not H.match(x)]
if not sigs: continue
m["kinds"] = sorted({kind(x) for x in sigs}); m["sigs"] = sigs
moves[m["run_key"]].append(m)
M = json.loads((B / "metrics.json").read_text())
runs = {}
for key, r in M["runs"].items():
ms = sorted(moves.get(key, []), key=lambda m: m["move"])
n = len(ms)
cnt = collections.Counter(k for m in ms for k in m["kinds"])
tot = sum(cnt[k] for k in KINDS) or 1
domains = {re.sub(r"^(webfetch|pdftext|doctext|browser:\w+|curl):", "", x) for m in ms for x in m["sigs"]
if re.match(r"^(webfetch|pdftext|doctext|browser:(asset|session|fetch|get|show)|curl):.+", x)}
searches = cnt["search"]; reads = cnt["page"] + cnt["official"]
runs[key] = dict(task=r["task"], model=r["model"], repeat=r["repeat"], acc=r.get("grade_accuracy"),
moves=n, **{f"share_{k}": cnt[k] / tot for k in KINDS},
domains=len(domains), reads_per_search=(reads / searches) if searches else None,
wall=r["wall_s"])
json.dump(runs, open(Path(sys.argv[2]) / "runs_profile.json", "w"), indent=0)
FEATS = [f"share_{k}" for k in KINDS] + ["domains", "moves"]
def mean(xs): xs = [x for x in xs if x is not None]; return st.mean(xs) if xs else None
models = ["P1", "DS", "GLM"]
def by_model(feat, rs):
return {m: mean([r[feat] for r in rs if r["model"] == m]) for m in models}
def spread(feat, rs):
v = by_model(feat, rs); xs = [x for x in v.values() if x is not None]
return max(xs) - min(xs)
def perm_p(feat, rs, n=5000, seed=0):
rng = random.Random(seed); obs = spread(feat, rs)
groups = collections.defaultdict(list)
for r in rs: groups[(r["task"], r["repeat"])].append(r)
hits = 0
for _ in range(n):
shuf = []
for g in groups.values():
labs = [r["model"] for r in g]; rng.shuffle(labs)
shuf += [dict(r, model=l) for r, l in zip(g, labs)]
if spread(feat, shuf) >= obs - 1e-12: hits += 1
return (hits + 1) / (n + 1)
rs = list(runs.values())
print("ALL TASKS (n=%d runs)" % len(rs))
res = {}
for f in FEATS:
v = by_model(f, rs); p = perm_p(f, rs)
res[f] = {"byModel": {m: round(x, 4) for m, x in v.items()}, "p": round(p, 4)}
print(f"{f:16s}", {m: round(x, 3) for m, x in v.items()}, "perm p =", round(p, 4))
json.dump(res, open(Path(sys.argv[2]) / "source_tests.json", "w"), indent=1)
print("\nPER TASK, share of search / page / data / process (P1|DS|GLM)")
for t in sorted({r['task'] for r in rs}):
sub = [r for r in rs if r["task"] == t]
row = []
for f in ["share_search", "share_page", "share_official", "share_data", "share_process", "domains"]:
v = by_model(f, sub); row.append(f"{f[6:] if f.startswith('share') else f}=" + "|".join(f"{v[m]:.2f}" if v[m] is not None else "-" for m in models))
print(t, " ".join(row))
|