Blame for build.py

"""Select foresight-swe-mini and reproduce every number on the card.

    python build.py verified.jsonl

verified.jsonl is data/verified.jsonl from shaswat123/Foresight-SWE. Writes
data/tasks.jsonl, data/leaderboard.csv and data/metrics.json. Upstream publishes
totals and per-model breakdowns only, so accuracy is simulated, not measured.
"""
import collections as C, csv, json, math, sys
import numpy as np
from scipy.stats import spearmanr

rows = [json.loads(l) for l in open(sys.argv[1])]
sizes = json.load(open("data/image_sizes.json"))
L = lambda v: json.loads(v) if isinstance(v, str) else (v or [])
N, K = len(rows), 15
col = lambda k: np.array([r[k] for r in rows])
ids, repo, lang, d1, d2, fm, tt, sec = (col(k) for k in
    ("instance_id", "repo", "language", "pr1_difficulty", "difficulty", "failure_mode", "pr1_task_type", "security"))
sev = np.array(["minor" if r["severity"] == "minor" else "major" for r in rows])
count = lambda k: np.array([len(L(r[k])) for r in rows])
p2p, f2p, ant = count("PASS_TO_PASS"), count("FAIL_TO_PASS"), count("ANTICIPATION_FAIL_TO_PASS")
patch_kb = np.array([len(r["patch"]) / 1e3 for r in rows])
tag = lambda r, k: sizes[r[k].split(":")[1]] / 1e9
image_gb = np.array([tag(r, "docker_image") for r in rows])
both_images_gb = image_gb + np.array([tag(r, "pr2base_image") for r in rows])

# Paper, Tables D.1 and D.2, in points:
# PR-1 easy/medium/hard, PR-2 easy/medium/hard, minor/major, loud/silent.
MODELS = ["Qwen3.8 Flash Next", "GLM-5.3 Flash", "DeepSeek-V4 Flash", "GPT-5.6 Sol", "GPT-5.6 Terra",
          "GPT-5.6 Luna", "Claude Opus 5", "Claude Sonnet 5", "Qwen3.5-397B-A17B", "Qwen3.8-27B", "Qwen3.5-27B"]
RESOLVE = np.array([
    [95.7, 97.4, 85.2, 91.2, 98.1, 85.7, 91.6, 92.2, 92.8, 91.5], [76.6, 80.5, 83.0, 73.6, 90.4, 91.4, 79.5, 81.4, 80.7, 80.6],
    [68.1, 59.7, 52.3, 55.2, 69.2, 54.3, 59.0, 58.1, 61.4, 56.6], [93.6, 81.8, 81.8, 80.0, 90.4, 91.4, 88.0, 82.2, 90.4, 80.6],
    [85.1, 79.2, 64.8, 76.8, 71.2, 71.4, 77.1, 72.9, 83.1, 69.0], [87.2, 76.6, 64.8, 72.0, 76.9, 77.1, 74.7, 73.6, 78.3, 71.3],
    [85.1, 74.0, 77.3, 74.4, 88.5, 74.3, 77.1, 78.3, 78.3, 77.5], [76.6, 57.1, 53.4, 56.8, 65.4, 62.9, 62.7, 58.1, 65.1, 56.6],
    [46.8, 23.4, 6.8, 22.4, 23.1, 17.1, 24.1, 20.2, 24.1, 20.2], [53.2, 27.3, 9.1, 25.6, 28.8, 20.0, 25.3, 25.6, 26.5, 24.8],
    [38.3, 23.4, 3.4, 17.6, 25.0, 11.4, 18.1, 18.6, 16.9, 19.4]]) / 100
FRR = np.array([
    [27.7, 31.2, 8.0, 28.8, 15.4, 0.0, 20.5, 20.9, 25.3, 17.8], [17.0, 14.3, 5.7, 12.0, 15.4, 2.9, 13.3, 10.1, 15.7, 8.5],
    [21.3, 29.9, 8.0, 21.6, 23.1, 2.9, 24.1, 15.5, 24.1, 15.5], [21.3, 23.4, 12.5, 24.8, 15.4, 0.0, 18.1, 18.6, 20.5, 17.1],
    [25.5, 16.9, 8.0, 20.8, 9.6, 2.9, 18.1, 13.2, 18.1, 13.2], [23.4, 23.4, 6.8, 20.8, 17.3, 0.0, 18.1, 15.5, 20.5, 14.0],
    [38.3, 26.0, 15.9, 28.0, 30.8, 2.9, 33.7, 18.6, 32.5, 19.4], [21.3, 20.8, 11.4, 19.2, 19.2, 5.7, 22.9, 13.2, 21.7, 14.0],
    [10.6, 11.7, 1.1, 8.8, 5.8, 2.9, 10.8, 4.7, 9.6, 5.4], [12.8, 10.4, 2.3, 8.8, 7.7, 2.9, 9.6, 6.2, 10.8, 5.4],
    [10.6, 9.1, 1.1, 7.2, 7.7, 0.0, 7.2, 5.4, 4.8, 7.0]]) / 100
FULL_FRR = np.array([20.75, 11.32, 18.87, 18.40, 15.09, 16.51, 24.53, 16.98, 7.08, 7.55, 6.13])
FULL_RESOLVE = np.array([91.98, 80.66, 58.49, 84.43, 74.53, 74.06, 77.83, 59.91, 21.70, 25.47, 18.40])

X = np.array([d1 == "easy", d1 == "medium", d1 == "hard", d2 == "easy", d2 == "medium", d2 == "hard",
              sev == "minor", sev == "major", fm == "loud", fm == "silent"], float).T
XX = np.c_[np.ones(N), X[:, [0, 1, 3, 4, 6, 8]]]
sig = lambda z: 1 / (1 + np.exp(-z))


def fit(margin):
    """Additive logit whose per-group means match one model's table row."""
    counts = margin * X.sum(0)
    t = np.r_[counts[:3].sum(), counts[[0, 1, 3, 4, 6, 8]]]
    b = np.zeros(XX.shape[1])
    b[0] = math.log((t[0] / N + 1e-3) / (1 - t[0] / N + 1e-3))
    for _ in range(200):
        p = sig(XX @ b)
        H = (XX * (p * (1 - p))[:, None]).T @ XX + 1e-3 * np.eye(len(b))
        b += 0.5 * np.linalg.solve(H, t - XX.T @ p)
    return b


bF = np.array([fit(m) for m in FRR])
pF, pR = sig(XX @ bF.T), sig(XX @ np.array([fit(m) for m in RESOLVE]).T)

feats = np.c_[X, *[(lang == l) * 1. for l in ("python", "javascript", "typescript", "go")],
              (tt == "feature") * 1., (tt == "bug") * 1., sec * 1.,
              (repo == "sveltejs/svelte") * 1., (repo == "django/django") * 1.]
Z = np.c_[np.log10(p2p + 1), np.log10(patch_kb + 1), np.log10(ant + 1)]


def objective(S):
    return (((feats[S].mean(0) - feats.mean(0)) ** 2).sum()
            + ((pF[S].mean(0) - pF.mean(0)) ** 2).sum() * 8 + ((pR[S].mean(0) - pR.mean(0)) ** 2).sum() * 2
            + (((Z[S].mean(0) - Z.mean(0)) / Z.std(0)) ** 2).sum() * .02)


def search(pool, seed, iters=25000):
    r = np.random.default_rng(seed)
    S = list(r.choice(pool, K, replace=False))
    over = lambda s: any(v > max(1, math.ceil((repo == rp).mean() * K)) for rp, v in C.Counter(repo[s]).items())
    v, T = objective(S) + over(S), .002
    for _ in range(iters):
        i, o = r.integers(K), r.choice(pool)
        if o in S:
            continue
        S2 = S.copy()
        S2[i] = o
        v2 = objective(S2) + over(S2)
        if v2 < v or r.random() < math.exp(-(v2 - v) / max(T, 1e-9)):
            S, v = S2, v2
        T *= .9995
    return sorted(S), v


# Keep runs short: leave out the tasks that are slow to grade.
pool = np.where((p2p <= 2000) & (ant <= 20) & (f2p <= 100) & (patch_kb <= 60) & (both_images_gb <= 3.5))[0]
S = min((search(pool, s) for s in range(8)), key=lambda x: x[1])[0]
# Removed after selection: upstream's PR-1 install recipe resolves numpy 2 before
# building pandas, so the build fails and no PR-1 test can run, even for the gold patch.
BROKEN = {"pandas__49324-49704"}
S = [i for i in S if ids[i] not in BROKEN]


def calibrate(lin, target):
    a, b = -8., 8.
    for _ in range(40):
        c = (a + b) / 2
        a, b = (c, b) if sig(lin + c).mean() < target else (a, c)
    return c


def simulate(sd, reps=500, seed=7):
    """Shared per-task difficulty (sd, logits) plus per-model noise, on top of the fit."""
    rng, out = np.random.default_rng(seed), C.defaultdict(list)
    for _ in range(reps):
        lin = XX @ bF.T + rng.normal(0, sd, N)[:, None] + rng.normal(0, .5, (N, 11))
        for m in range(11):
            lin[:, m] += calibrate(lin[:, m], FULL_FRR[m] / 100)
        Y = (rng.random((N, 11)) < sig(lin)) * 1.
        full = Y.mean(0)
        designs = {f"random {k}": rng.choice(N, k, replace=False) for k in (10, len(S), 20, 30, 60, 100)}
        designs["foresight-swe-mini"] = np.array(S)
        for name, D in designs.items():
            e = Y[D].mean(0)
            out[name, "error_points"].append(np.abs(e - full).mean() * 100)
            for gap in (.05, .10):
                pairs = [(a, b) for a in range(11) for b in range(a + 1, 11) if abs(full[a] - full[b]) >= gap]
                out[name, f"pairs_{int(gap * 100)}pt"].append(
                    sum((e[a] - e[b]) * (full[a] - full[b]) > 0 for a, b in pairs) / max(len(pairs), 1) * 100)
    table = C.defaultdict(dict)
    for (name, metric), v in out.items():
        table[name][metric] = round(float(np.mean(v)), 1)
    return table


keep = ("instance_id", "repo", "language", "earlier_pr", "later_pr", "pr1_difficulty", "difficulty", "severity",
        "failure_mode", "pr1_task_type", "security", "docker_image")
with open("data/tasks.jsonl", "w") as f:
    for i in S:
        row = {k: rows[i][k] for k in keep}
        row.update(n_fail_to_pass=int(f2p[i]), n_anticipation=int(ant[i]), n_pass_to_pass=int(p2p[i]),
                   image_gb=round(float(image_gb[i]), 2))
        f.write(json.dumps(row) + "\n")
with open("data/leaderboard.csv", "w", newline="") as f:
    w = csv.writer(f)
    w.writerow(["model", "resolve_all_212", "frr_all_212", "predicted_frr_on_mini"])
    for m, name in enumerate(MODELS):
        w.writerow([name, FULL_RESOLVE[m], FULL_FRR[m], round(float(pF[S].mean(0)[m] * 100), 1)])
mix = dict(language=lang, pr1_difficulty=d1, pr2_difficulty=d2, severity=sev, detectability=fm, task_type=tt, security=sec)
metrics = dict(
    n_pool=int(len(pool)), image_gb=round(float(image_gb[S].sum()), 1), max_pass_to_pass=int(p2p[S].max()),
    repos=len(set(repo[S])),
    predicted_frr_max_error=round(float(np.abs(pF[S].mean(0) - pF.mean(0)).max() * 100), 1),
    predicted_resolve_max_error=round(float(np.abs(pR[S].mean(0) - pR.mean(0)).max() * 100), 1),
    variance_explained_frr=[round(float(x), 3) for x in pF.var(0) / (pF.mean(0) * (1 - pF.mean(0)))],
    spearman_resolve_frr_all=round(float(spearmanr(FULL_RESOLVE, FULL_FRR)[0]), 2),
    mix={k: dict(mini=dict(C.Counter(str(x) for x in v[S])),
                 proportional={str(a): round(b / N * len(S), 1) for a, b in C.Counter(v.tolist()).items()})
         for k, v in mix.items()},
    simulation={str(sd): simulate(sd) for sd in (.5, 1, 1.5)})
json.dump(metrics, open("data/metrics.json", "w"), indent=1)
print(" ".join(str(ids[i]) for i in S))