Datasets:
File size: 4,819 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 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 | #!/usr/bin/env python3
"""Build leakage-controlled splits of the CircuitQA corpus.
S1 sample-level (random split by sample; a design may appear in several partitions)
S2 design-disjoint (no source_file shared between train and test)
S3 repo-disjoint (no upstream repository shared between train and test)
Also emits a frozen stratified n=500 evaluation sample from the S3 test set,
used for BOTH CircuitQA and every baseline model so numbers are comparable.
"""
import json, collections, random, os
SRC = "training/data/v7_clean"
OUT = "revision/splits"
SEED = 42
CATS = ["structural_analysis","design_quality","functional_behavior",
"code_understanding","optimization","cross_stage"]
def repo_of(sf):
p = (sf or "").split("/")
return "/".join(p[:2]) if len(p) > 1 else (sf or "unknown")
rows = []
for split in ("train","val","test"):
with open(f"{SRC}/{split}.jsonl") as f:
for line in f:
d = json.loads(line)
d["_repo"] = repo_of(d.get("source_file"))
rows.append(d)
print(f"loaded {len(rows)} samples")
by_repo = collections.defaultdict(list)
for d in rows: by_repo[d["_repo"]].append(d)
sizes = {r: len(v) for r, v in by_repo.items()}
total = len(rows)
# ---- S3: repo-disjoint. Prefer several mid-sized repos over one large one,
# so the held-out set spans multiple design styles rather than one vendor's.
rng = random.Random(SEED)
ranked = sorted(sizes, key=lambda r: -sizes[r])
mid = [r for r in ranked if 300 <= sizes[r] <= 3000] # candidate pool
rng.shuffle(mid)
def take(pool, target, used):
sel, n = [], 0
for r in pool:
if r in used: continue
if n >= target * 0.90: break
if n + sizes[r] > target * 1.20: continue
sel.append(r); n += sizes[r]
return sel, n
used = set()
test_repos, n_te = take(mid, total*0.10, used); used |= set(test_repos)
val_repos, n_va = take(mid, total*0.10, used); used |= set(val_repos)
train_repos = [r for r in ranked if r not in used]
s3 = {"train": [d for r in train_repos for d in by_repo[r]],
"val": [d for r in val_repos for d in by_repo[r]],
"test": [d for r in test_repos for d in by_repo[r]]}
# ---- S2: design-disjoint. Hold out whole source_files, stratified by category.
files = sorted({d["source_file"] for d in rows})
rng2 = random.Random(SEED); rng2.shuffle(files)
n_f = len(files)
te_f = set(files[:int(n_f*0.10)]); va_f = set(files[int(n_f*0.10):int(n_f*0.20)])
s2 = {"train": [], "val": [], "test": []}
for d in rows:
k = "test" if d["source_file"] in te_f else "val" if d["source_file"] in va_f else "train"
s2[k].append(d)
def report(name, s):
print(f"\n=== {name} ===")
for k in ("train","val","test"):
cc = collections.Counter(d["category"] for d in s[k])
nf = len({d["source_file"] for d in s[k]})
print(f" {k:5s} n={len(s[k]):6d} ({100*len(s[k])/total:4.1f}%) designs={nf:5d} cats={len(cc)}/6")
tr_r = {d["_repo"] for d in s["train"]}; te_r = {d["_repo"] for d in s["test"]}
tr_f = {d["source_file"] for d in s["train"]}; te_f2 = {d["source_file"] for d in s["test"]}
tr_m = {d["module_name"] for d in s["train"]}; te_m = {d["module_name"] for d in s["test"]}
print(f" LEAK repo overlap : {len(tr_r & te_r)}")
print(f" LEAK design overlap : {len(tr_f & te_f2)}")
print(f" LEAK module overlap : {len(tr_m & te_m)} / {len(te_m)}")
report("S2 design-disjoint", s2)
report("S3 repo-disjoint", s3)
os.makedirs(OUT, exist_ok=True)
for tag, s in (("s2_design_disjoint", s2), ("s3_repo_disjoint", s3)):
for k in ("train","val","test"):
with open(f"{OUT}/{tag}.{k}.jsonl","w") as f:
for d in s[k]:
d = {kk: vv for kk, vv in d.items() if kk != "_repo"}
f.write(json.dumps(d) + "\n")
# ---- frozen stratified n=500 evaluation sample from S3 test
rng3 = random.Random(SEED)
pool = collections.defaultdict(list)
for d in s3["test"]: pool[d["category"]].append(d)
present = [c for c in CATS if pool[c]]
FLOOR, N = 50, 500
alloc = {c: FLOOR for c in present}
rem = N - FLOOR*len(present)
tot_rest = sum(len(pool[c]) for c in present)
for c in present:
alloc[c] += int(rem * len(pool[c]) / tot_rest)
sample = []
for c in present:
k = min(alloc[c], len(pool[c]))
sample += rng3.sample(pool[c], k)
rng3.shuffle(sample)
with open(f"{OUT}/eval_sample_n{len(sample)}.jsonl","w") as f:
for d in sample:
d = {kk: vv for kk, vv in d.items() if kk != "_repo"}
f.write(json.dumps(d) + "\n")
cc = collections.Counter(d["category"] for d in sample)
print(f"\n=== frozen eval sample: n={len(sample)} ===")
for c in CATS:
if cc[c]: print(f" {c:22s} {cc[c]:4d}")
print(f"\nheld-out TEST repos (S3): {test_repos}")
print(f"held-out VAL repos (S3): {val_repos}")
|