Blame for scripts/build_clean_set.py

"""Rebuilds data/clean-16k.jsonl (as pairs) from the training corpus SagheerLab/Arithmetic-Reasoning.
   python3 scripts/build_clean_set.py > build.out          (about 10 minutes, downloads the 5 GB corpus shard by shard and deletes each; CPU only, no GPU, needs pyarrow + huggingface_hub)
Prints  PAIRS_JSON {type: [[a, b], ...]}  (the operands of the 2,000 clean questions per type, in file order; the answer is a op b)  and
        LEGACY_FLAGS_JSON [ids of the legacy-14k items that occur in the corpus].
A question is 'seen' when `a op b` (or `b op a` for + and *) occurs in the corpus as an expression in `instruction` or as a step `a op b = c` in `reasoning`."""
import os, re, sys, time, json, collections, subprocess
subprocess.run([sys.executable, "-m", "pip", "install", "-q", "pyarrow", "huggingface_hub"], check=False)
import numpy as np, pyarrow.parquet as pq
from huggingface_hub import list_repo_files, hf_hub_download
REPO = "SagheerLab/Arithmetic-Reasoning"
NORM = lambda s: s.replace("×", "*").replace("x", "*").replace("÷", "/")
P_INST = re.compile(r"\d+(?:\s*[+\-*×x/÷]\s*\d+)+")
P_STEP = re.compile(r"(\d+)\s*([+\-*×x/÷])\s*(\d+)\s*=")
P_ONE = re.compile(r"(\d+)([+\-*/])(\d+)")
seen = set(); n = 0; t0 = time.time()
files = sorted(f for f in list_repo_files(REPO, repo_type="dataset") if f.endswith(".parquet"))
cols = None
for fi, f in enumerate(files):
    p = hf_hub_download(REPO, f, repo_type="dataset"); pf = pq.ParquetFile(p)
    cols = [c for c in ("instruction", "reasoning") if c in pf.schema_arrow.names]
    for batch in pf.iter_batches(batch_size=20000, columns=cols):
        arrs = [c.to_pylist() for c in batch.columns]
        for row in zip(*arrs):
            n += 1
            q = row[0] or ""
            for e in P_INST.findall(q):
                m = P_ONE.fullmatch(re.sub(r"\s+", "", NORM(e)))
                if m: seen.add((int(m.group(1)), m.group(2), int(m.group(3))))
            if len(row) > 1 and row[1]:
                for a, o, b in P_STEP.findall(row[1]):
                    seen.add((int(a), NORM(o), int(b)))
    os.remove(p)
    print(f"[scan] shard {fi+1}/{len(files)} rows {n:,} distinct single-op expressions {len(seen):,} cols {cols} {time.time()-t0:.0f}s", flush=True)
cov = collections.Counter((o, len(str(a)), len(str(b))) for a, o, b in seen)
print("COVERAGE_JSON", json.dumps({f"{o} {da}x{db}": c for (o, da, db), c in sorted(cov.items()) if c > 500}), flush=True)

def present(a, o, b): return (a, o, b) in seen or (o in "+*" and (b, o, a) in seen)

# type -> (digits of a, digits of b, operator, seed, keep the original legacy items first)
TYPES = {"3ds": (3, 3, "-", 104, True), "4da": (4, 4, "+", 105, True), "5da": (5, 5, "+", 106, True),
         "5ds": (5, 5, "-", 108, False), "4dm": (4, 4, "*", 110, False), "5dm": (5, 5, "*", 111, False), "34m": (3, 4, "*", 113, False)}
INFEASIBLE_PROBE = {"3da": (3, 3, "+", 103, True), "3dm": (3, 3, "*", 109, False)}      # probed only: the training data covers (almost) the whole space
DIV = ((11, 99), (101, 999))                                                                # divisor 2-digit, quotient 3-digit, dividend = their product (1,111 .. 98,901)

def draw(rng, da, db, op):
    a = int(rng.randint(10 ** (da - 1), 10 ** da)); b = int(rng.randint(10 ** (db - 1), 10 ** db))
    if op == "-" and b > a: a, b = b, a
    return a, b

def pick(tag, da, db, op, seed, legacy, cap=2_000_000):
    out, got, stats = [], set(), collections.Counter()
    def take(a, b, src):
        if (a, b) in got or (op in "+*" and (b, a) in got): stats["dup"] += 1; return
        if present(a, op, b): stats["in_training_" + src] += 1; return
        got.add((a, b)); out.append([a, b]); stats["kept_" + src] += 1
    if legacy:
        rng = np.random.RandomState(seed)
        for _ in range(2000): take(*draw(rng, da, db, op), "legacy")
    rng = np.random.RandomState(seed + 1000); tries = 0
    while len(out) < 2000 and tries < cap:
        tries += 1; take(*draw(rng, da, db, op), "fresh")
    print(f"[pick] {tag}: {len(out)} items {dict(stats)} draws {tries}", flush=True)
    return out

res = {}
for tag, (da, db, op, sd, lg) in TYPES.items(): res[tag] = pick(tag, da, db, op, sd, lg)
for tag, (da, db, op, sd, lg) in INFEASIBLE_PROBE.items(): pick(tag, da, db, op, sd, lg, cap=300_000)
(dl, dh), (ql, qh) = DIV
rng = np.random.RandomState(112 + 1000); out, got, hit = [], set(), 0
while len(out) < 2000:
    y, q = int(rng.randint(dl, dh + 1)), int(rng.randint(ql, qh + 1)); x = y * q
    if (x, y) in got: continue
    if present(x, "/", y): hit += 1; continue
    got.add((x, y)); out.append([x, y])
res["div"] = out; print(f"[div] 2000 items, rejected {hit} that were in the training data", flush=True)
for t, v in res.items():
    assert len(v) == 2000 and len({tuple(p) for p in v}) == 2000, t
    op = "/" if t == "div" else TYPES[t][2]
    assert not any(present(a, op, b) for a, b in v), t
# legacy 14k per-item flags (for the archive file)
LEGACY_ALL = {"2da": (2, "+", 101), "2ds": (2, "-", 102), "3da": (3, "+", 103), "3ds": (3, "-", 104), "4da": (4, "+", 105), "5da": (5, "+", 106), "2dm": (2, "*", 107)}
flags = []
for tag, (d, op, sd) in LEGACY_ALL.items():
    rng = np.random.RandomState(sd)
    for k in range(2000):
        a, b = draw(rng, d, d, op)
        if present(a, op, b): flags.append(f"{tag}-{k:04d}")
print("RESULT_JSON", json.dumps({"rows": n, "legacy_flagged": len(flags), "legacy_flagged_by_type": dict(collections.Counter(i.split('-')[0] for i in flags)), "sizes": {k: len(v) for k, v in res.items()}}), flush=True)
print("PAIRS_JSON", json.dumps(res), flush=True)
print("LEGACY_FLAGS_JSON", json.dumps(flags), flush=True)