vh
739aa03123
Latency, measured from nh3-dev (3 runs x 20 per condition; network floor 31 ms): - /decide short: 71 ms end to end, 38 ms server-side; - /decide with a ~2,000-token state: 210 / 169 ms; - shared, 3 rotations: 113 / 79 ms; - shared, 6 orderings: 137 / 99 ms. Qwen3.5's fast kernels (causal_conv1d, flash-linear-attention) are not installed, so transformers falls back to its reference PyTorch paths. That is a speed lever, and using it needs a parity re-check. Averaging over option orderings, on SemIf authored144 + perturbations108 (252 rows, 72 groups): - a single ordering scores 78.6%; - log-mean over the 3 rotations scores 87.7% (+9.1 pts, group-bootstrap 95% CI +4.7 to +13.8); - all 6 permutations score 88.1%. Rotations capture nearly all of the gain. Rows where the rotations agree unanimously (161) are 94.4% accurate; split rows (91) are 75.8%.
90 lines
4.7 KiB
Python
90 lines
4.7 KiB
Python
"""SPIKE (2026-09-27): does averaging over option orderings help, and is agreement a useful
|
|
ambiguity signal? Throwaway measurement, no service change: every labelled row goes out as ONE
|
|
/decide/shared request carrying all 3! = 6 orderings of its options.
|
|
|
|
Data: SemIf authored144 + perturbations108 (labelled, 3 options each), pinned commit.
|
|
Conditions derived from the same 6 scored orderings per row:
|
|
single-original the caller's own order (what the service does today)
|
|
single-expected mean accuracy over the 6 single orderings (a caller's expected luck)
|
|
rotations the 3 cyclic shifts of the original order, log-mean combined
|
|
all-6 all 6 permutations, log-mean combined
|
|
Paired group-bootstrap (resampling group_id, 10k) for each combined condition minus
|
|
single-original, so the delta is judged against its own sampling noise.
|
|
SEMIF_DIR=... SEMIF_URL=... SEMIF_TOKEN=... uv run --with httpx python averaging_spike.py out.json
|
|
"""
|
|
import itertools, json, math, os, random, statistics as st, sys, time
|
|
from pathlib import Path
|
|
import httpx
|
|
|
|
S, U = Path(os.environ["SEMIF_DIR"]), os.environ["SEMIF_URL"]
|
|
H = {"Authorization": f"Bearer {os.environ['SEMIF_TOKEN']}"}
|
|
rows = []
|
|
for name in ("authored144", "perturbations108"):
|
|
rows += [dict(json.loads(l), set=name) for l in (S / f"benchmarks/data/{name}.jsonl").read_text().splitlines() if l.strip()]
|
|
|
|
|
|
def logmean(dists): # dists: list of {option_id: p}; mean log p per option, renormalised
|
|
ids = dists[0].keys()
|
|
m = {k: st.fmean(math.log(max(d[k], 1e-12)) for d in dists) for k in ids}
|
|
top = max(m.values())
|
|
z = sum(math.exp(v - top) for v in m.values())
|
|
return {k: math.exp(v - top) / z for k, v in m.items()}
|
|
|
|
|
|
argmax = lambda d: max(d, key=d.get)
|
|
out, t0 = [], time.time()
|
|
for r in rows:
|
|
opts = r["options"]
|
|
perms = list(itertools.permutations(range(len(opts))))
|
|
body = {"state": r["state"], "decisions": [
|
|
{"id": f"p{i}", "question": r["question"], "options": [opts[j] for j in p]} for i, p in enumerate(perms)]}
|
|
res = httpx.post(f"{U}/decide/shared", headers=H, json=body, timeout=120)
|
|
res.raise_for_status()
|
|
dists = [dict(zip(x["option_ids"], x["probabilities"])) for x in res.json()["results"]]
|
|
by_perm = dict(zip(perms, dists))
|
|
ident = tuple(range(len(opts)))
|
|
rot = [tuple((k + i) % len(opts) for k in ident) for i in range(len(opts))]
|
|
gold = opts[r["label"]]["id"]
|
|
rot_d, all_d = logmean([by_perm[p] for p in rot]), logmean(dists)
|
|
out.append({
|
|
"id": r["id"], "set": r["set"], "group": r["group_id"], "gold": gold,
|
|
"single_original": argmax(by_perm[ident]) == gold,
|
|
"single_expected": st.fmean(argmax(d) == gold for d in dists),
|
|
"rotations": argmax(rot_d) == gold,
|
|
"all6": argmax(all_d) == gold,
|
|
"agree_rot": sum(argmax(by_perm[p]) == argmax(rot_d) for p in rot) / len(rot),
|
|
"agree_all6": sum(argmax(d) == argmax(all_d) for d in dists) / len(dists),
|
|
"first_position_wins": sum(argmax(d) == opts[p[0]]["id"] for p, d in by_perm.items()) / len(perms),
|
|
})
|
|
wall = time.time() - t0
|
|
|
|
|
|
def boot(key, reps=10000, seed=7):
|
|
groups = {}
|
|
for o in out:
|
|
groups.setdefault(o["group"], []).append(o)
|
|
keys, rng, deltas = list(groups), random.Random(seed), []
|
|
for _ in range(reps):
|
|
sample = [o for g in (rng.choice(keys) for _ in keys) for o in groups[g]]
|
|
deltas.append(st.fmean(o[key] for o in sample) - st.fmean(o["single_original"] for o in sample))
|
|
deltas.sort()
|
|
return round(deltas[int(0.025 * reps)], 4), round(deltas[int(0.975 * reps)], 4)
|
|
|
|
|
|
acc = lambda key, sub=out: round(st.fmean(o[key] for o in sub), 4)
|
|
report = {"rows": len(out), "groups": len({o['group'] for o in out}), "wall_s": round(wall, 1),
|
|
"accuracy": {k: acc(k) for k in ("single_original", "single_expected", "rotations", "all6")},
|
|
"delta_vs_single_original_95ci": {k: boot(k) for k in ("rotations", "all6")},
|
|
"first_position_win_rate_mean": acc("first_position_wins"),
|
|
"by_set": {s: {k: acc(k, [o for o in out if o["set"] == s]) for k in ("single_original", "rotations", "all6")}
|
|
for s in ("authored144", "perturbations108")}}
|
|
for key in ("agree_rot", "agree_all6"):
|
|
unan = [o for o in out if o[key] == 1.0]
|
|
split = [o for o in out if o[key] < 1.0]
|
|
cond = "rotations" if key == "agree_rot" else "all6"
|
|
report[f"{key}: accuracy when unanimous vs split"] = {
|
|
"unanimous": {"rows": len(unan), "accuracy": acc(cond, unan) if unan else None},
|
|
"split": {"rows": len(split), "accuracy": acc(cond, split) if split else None}}
|
|
json.dump({"report": report, "rows": out}, open(sys.argv[1], "w"), indent=1)
|
|
print(json.dumps(report, indent=1))
|