Datasets:
File size: 12,592 Bytes
8ad880b | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 | """Build the SeqRoute-1M master query pool (30,000), the 20,000-query generation subset, the 10,000 expansion pool, splits and nested scaling subsets.
Usage: python build_queries.py
Writes /mnt/localssd/seqroute1m/queries/{master_queries.parquet, queries_20k.parquet} and manifests/{queries,splits,subsets}*.json
Everything is deterministic from SEED. Splits are query-level (grouped by source group_id so paraphrase / same-passage / same-table / same-MBPP-task items never straddle splits),
stratified by task family x source dataset, 70/15/15, and are defined over the whole 30k master; the 20k subset inherits them.
"""
import hashlib
import json
import random
import sys
import time
from collections import defaultdict
from pathlib import Path
import pandas as pd
sys.path.insert(0, "src")
sys.path.insert(0, ".")
from llmroute.seqroute import families as F, synth as S
import os
ROOT = Path(os.environ.get("SEQROUTE_BUILD_ROOT", "/mnt/localssd/seqroute1m"))
SEED = 20261010
N_MASTER, N_GEN = 30000, 20000
# (family) -> (base target in master, max_tokens non-reasoning, max_tokens reasoning)
FAMILIES = {
"math_word": (1500, 1536, 6144), "math_competition": (1700, 3072, 8192), "knowledge": (1700, 1024, 4096), "science": (1700, 1536, 6144), "logic_reasoning": (1700, 1536, 6144),
"code_generation": (1700, 2048, 6144), "code_repair": (1300, 2048, 6144), "summ_news": (1500, 384, 4096), "summ_abstractive": (1500, 256, 4096), "simplification": (1500, 256, 3072),
"translation": (1500, 512, 4096), "information_extraction": (1500, 768, 4096), "classification": (1500, 64, 3072), "json_generation": (1200, 512, 3072),
"instruction_following": (1600, 2048, 6144), "long_context_qa": (1100, 512, 4096), "reading_comprehension": (1500, 256, 4096), "table_reasoning": (1500, 1024, 4096),
"planning": (1400, 4096, 8192), "multilingual": (1500, 1536, 6144),
}
CTX = 16384
TOK_SAFETY = 1.2 # prompt tokens are counted with the Qwen3 tokenizer; other models' tokenizers differ by up to ~15%
def hh(*parts, n=16):
return hashlib.sha1("\x1f".join(str(p) for p in parts).encode()).hexdigest()[:n]
def waterfill(avail: dict, total: int) -> dict:
"""Equal allocation across sources, capped by availability, remainder redistributed."""
alloc = {k: 0 for k in avail}
left, active = total, [k for k in avail if avail[k] > 0]
while left > 0 and active:
share = max(left // len(active), 1)
nxt = []
for k in active:
take = min(share, avail[k] - alloc[k], left)
alloc[k] += take
left -= take
if alloc[k] < avail[k]:
nxt.append(k)
if all(min(share, avail[k] - alloc[k]) == 0 for k in nxt) and not nxt:
break
active = nxt
return alloc
def pick(rows: list[dict], total: int, rng: random.Random) -> list[dict]:
by = defaultdict(list)
for r in rows:
by[r["source_dataset"]].append(r)
for v in by.values():
rng.shuffle(v)
al = waterfill({k: len(v) for k, v in by.items()}, min(total, len(rows)))
out = []
for k, v in by.items():
out += v[:al[k]]
return out
def assign_splits(df):
df["split"] = ""
for (fam, src), g in df.groupby(["task_family", "source_dataset"]):
groups = sorted(g.group_id.unique(), key=lambda x: hh("split", SEED, x))
sizes = g.group_id.value_counts()
n, acc, assign = len(g), 0, {}
for gid in groups:
frac = acc / n
assign[gid] = "train" if frac < 0.70 else ("val" if frac < 0.85 else "test")
acc += sizes[gid]
df.loc[g.index, "split"] = g.group_id.map(assign)
df["legacy_split"] = df["split"]
nstrata = df.groupby("group_id").apply(lambda g: g[["task_family", "source_dataset"]].drop_duplicates().shape[0])
linked = set(nstrata[nstrata > 1].index)
fixed = {}
for gid in linked:
u = int(hh("linked", SEED, gid, n=8), 16) / 16 ** 8
fixed[gid] = "train" if u < 0.70 else ("val" if u < 0.85 else "test")
for (fam, src), g in df.groupby(["task_family", "source_dataset"]):
sizes = g.group_id.value_counts()
cnt = {"train": 0, "val": 0, "test": 0}
assign = {}
for gid in sizes.index:
if gid in fixed:
assign[gid] = fixed[gid]; cnt[fixed[gid]] += sizes[gid]
n, tot = len(g), 0
for gid in sorted((x for x in sizes.index if x not in fixed), key=lambda x: hh("split", SEED, x)):
tot_now = sum(cnt.values()) or 1
deficit = {sp: w * max(n, 1) - cnt[sp] for sp, w in (("train", 0.70), ("val", 0.15), ("test", 0.15))}
sp = max(deficit, key=lambda k: (deficit[k], k == "train"))
assign[gid] = sp; cnt[sp] += sizes[gid]
df.loc[g.index, "split"] = g.group_id.map(assign)
return df
def main():
t0 = time.time()
fam_fns = {"math_word": F.math_word, "math_competition": F.math_competition, "knowledge": F.knowledge, "science": F.science, "logic_reasoning": F.logic_reasoning,
"code_generation": F.code_generation, "summ_news": F.summ_news, "summ_abstractive": F.summ_abstractive, "simplification": F.simplification, "translation": F.translation,
"information_extraction": F.information_extraction, "classification": F.classification, "json_generation": F.json_generation, "instruction_following": S.instruction_following,
"long_context_qa": F.long_context_qa, "reading_comprehension": F.reading_comprehension, "table_reasoning": F.table_reasoning, "planning": S.planning, "multilingual": F.multilingual}
pools = {}
for fam, fn in fam_fns.items():
pools[fam] = fn()
print(fam, len(pools[fam]), f"{time.time() - t0:.0f}s", flush=True)
pools["code_repair"] = F.code_repair_humanevalfix() + F.code_repair_mbpp(1300, SEED) + F.code_repair_mbpp(1300, SEED + 1)
print("code_repair", len(pools["code_repair"]), f"{time.time() - t0:.0f}s", flush=True)
from transformers import AutoTokenizer
tok = AutoTokenizer.from_pretrained("/mnt/localssd/models/qwen3_8b")
rng = random.Random(SEED)
chosen, dropped = {}, defaultdict(int)
for fam, rows in pools.items():
seen, uniq = set(), []
for r in rows:
k = hh(fam, r["prompt"])
if k in seen:
dropped[fam, "duplicate_prompt"] += 1
continue
seen.add(k)
uniq.append(r)
base, _, rmax = FAMILIES[fam]
cand = pick(uniq, int(base * 1.35) + 50, rng)
lens = [len(x) for x in tok([r["prompt"] for r in cand], add_special_tokens=False)["input_ids"]]
ok = []
for r, n in zip(cand, lens):
r["prompt_tokens_qwen3"] = n
if n * TOK_SAFETY + rmax + 256 > CTX:
dropped[fam, "too_long"] += 1
continue
ok.append(r)
chosen[fam] = ok
print(fam, "candidates", len(cand), "after length filter", len(ok), flush=True)
# sizes: base targets capped by availability; the shortfall is redistributed (proportionally to base target) to families with spare candidates
tgt = {f: min(FAMILIES[f][0], len(chosen[f])) for f in chosen}
while sum(tgt.values()) < N_MASTER:
spare = {f: len(chosen[f]) - tgt[f] for f in chosen if len(chosen[f]) > tgt[f]}
if not spare:
raise SystemExit(f"cannot reach {N_MASTER}; have {sum(tgt.values())}")
need = N_MASTER - sum(tgt.values())
wsum = sum(FAMILIES[f][0] for f in spare)
for f in spare:
tgt[f] += min(spare[f], max(1, round(need * FAMILIES[f][0] / wsum)))
if sum(tgt.values()) >= N_MASTER:
break
while sum(tgt.values()) > N_MASTER:
f = max(tgt, key=tgt.get)
tgt[f] -= 1
master = []
for fam, rows in chosen.items():
sel = pick(rows, tgt[fam], rng)
for r in sel:
r["max_tokens_nonreasoning"], r["max_tokens_reasoning"] = FAMILIES[fam][1], FAMILIES[fam][2]
master += sel
df = pd.DataFrame(master)
df["query_id"] = [f"{f[:4]}_{hh(f, s, i, p, n=12)}" for f, s, i, p in zip(df.task_family, df.source_dataset, df.source_id, df.prompt)]
assert df.query_id.is_unique, "query_id collision"
df["payload"] = df.payload.map(lambda p: json.dumps(p, ensure_ascii=False, default=lambda o: o.tolist() if hasattr(o, "tolist") else str(o)))
print("master", len(df), dict(df.task_family.value_counts()), flush=True)
# ---- splits over the 30k master: 70/15/15 by group within (family, source) ----------------------------------------------------------
# `legacy_split` (strata handled independently) is only used to draw the 20k subset; the final `split` additionally keeps a group that spans several
# strata (e.g. HumanEval / HumanEval-Fix, MBPP / MBPP-mutants) in a single split.
df = assign_splits(df)
# ---- 20k generation subset: proportional within (family, source, split) ---------------------------------------------------------------
df["in_20k"] = False
r2 = random.Random(SEED + 7)
for key, g in df.groupby(["task_family", "source_dataset", "legacy_split"]):
k = round(len(g) * N_GEN / N_MASTER)
idx = list(g.index)
r2.shuffle(idx)
df.loc[idx[:k], "in_20k"] = True
# fix rounding to exactly 20,000
cur = int(df.in_20k.sum())
pool_in, pool_out = list(df.index[df.in_20k]), list(df.index[~df.in_20k])
r2.shuffle(pool_in); r2.shuffle(pool_out)
if cur > N_GEN:
df.loc[pool_in[:cur - N_GEN], "in_20k"] = False
elif cur < N_GEN:
df.loc[pool_out[:N_GEN - cur], "in_20k"] = True
assert int(df.in_20k.sum()) == N_GEN
# ---- nested scaling subsets within the 20k: stratified interleaved ordering over the train split ----------------------------------
df["scale_rank"] = -1.0
tr = df[(df.split == "train") & df.in_20k]
for _, g in tr.groupby(["task_family", "source_dataset"]):
idx = sorted(g.index, key=lambda i: hh("scale", SEED, df.at[i, "query_id"]))
for rank, i in enumerate(idx):
df.at[i, "scale_rank"] = (rank + 0.5) / len(idx)
order = tr.index[tr.index.map(lambda i: True)]
order = sorted(order, key=lambda i: (df.at[i, "scale_rank"], hh("tie", df.at[i, "query_id"])))
df["scale_order"] = -1
for pos, i in enumerate(order):
df.at[i, "scale_order"] = pos
df["prompt_chars"] = df.prompt.str.len()
(ROOT / "queries").mkdir(parents=True, exist_ok=True)
df.to_parquet(ROOT / "queries/master_queries.parquet")
df[df.in_20k].reset_index(drop=True).to_parquet(ROOT / "queries/queries_20k.parquet")
M = ROOT / "manifests"
M.mkdir(exist_ok=True)
for sp in ("train", "val", "test"):
json.dump(sorted(df[df.split == sp].query_id), open(M / f"master_{sp}_ids.json", "w"))
json.dump(sorted(df[(df.split == sp) & df.in_20k].query_id), open(M / f"gen20k_{sp}_ids.json", "w"))
json.dump(sorted(df[~df.in_20k].query_id), open(M / "expansion_10k_ids.json", "w"))
trn = df[(df.split == "train") & df.in_20k].sort_values("scale_order")
for n in (1000, 2000, 5000, 10000, len(trn)):
json.dump(list(trn.query_id[:n]), open(M / f"scaling_train_{n}_ids.json", "w"))
man = dict(created=time.strftime("%F %T"), seed=SEED, n_master=len(df), n_gen20k=int(df.in_20k.sum()), n_expansion=int((~df.in_20k).sum()),
per_family_master=df.groupby("task_family").size().to_dict(), per_family_20k=df[df.in_20k].groupby("task_family").size().to_dict(),
per_source_20k=df[df.in_20k].groupby(["task_family", "source_dataset"]).size().rename("n").reset_index().to_dict("records"),
split_master=df.split.value_counts().to_dict(), split_20k=df[df.in_20k].split.value_counts().to_dict(),
dropped={f"{k[0]}:{k[1]}": v for k, v in dropped.items()}, token_safety=TOK_SAFETY, ctx=CTX,
split_rule="group-level (group_id) 70/15/15 within (task_family, source_dataset) over the 30k master; 20k subset = proportional sample within (family, source, split)",
scaling_note="scaling_train_N = first N train queries of the 20k set in a stratified interleaved order (nested); val/test fixed. Train of the 20k is 14,000; sizes beyond need the expansion pool.")
json.dump(man, open(M / "queries_manifest.json", "w"), indent=1)
print(df[df.in_20k].groupby(["task_family", "split"]).size().unstack().to_string())
print("done", f"{time.time() - t0:.0f}s")
if __name__ == "__main__":
main()
|