File size: 14,200 Bytes
1a6af84 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 | """3-way split (train_proj / train_rerank / val) — v2.
Improvements over v1:
1. Softer cutoff (default 2017-01-01) — gives more candidate sessions for
the holdout while still preserving the temporal property:
no train_proj session is dated AFTER the latest val session.
2. Stratification by user warmth (warm/cold in cf-bpr): the simulated-cold
rate is computed among WARM users only, so the holdout's real-cold rate
matches dev's 25.8% target rather than 26% of everyone.
3. Stratification by dominant (category, specificity): inside the holdout
pool, users are bucketed by their most frequent goal pair and sampled
proportionally so train_rerank and val have similar category mixes.
4. Quota-based selection rather than greedy: each stratum gets an integer
session quota, and users are picked deterministically by hash within
each stratum until the quota is met.
Output: models/splits/train_val.parquet with columns
session_id, user_id, split, simulate_cold
Run:
uv run python -m scripts.launchers.make_train_val_split_v2 \\
--cutoff-date 2017-01-01
"""
from __future__ import annotations
import argparse
import hashlib
from collections import defaultdict, Counter
from pathlib import Path
import polars as pl
#DATA = Path("./data/talkpl-ai")
# TRAIN_PARQUET = DATA / "TalkPlayData-Challenge-Dataset/data/train-00000-of-00001.parquet"
# DEV_PARQUET = DATA / "TalkPlayData-Challenge-Dataset/data/test-00000-of-00001.parquet"
# USER_EMB_TRAIN = DATA / "TalkPlayData-Challenge-User-Embeddings/data/train-00000-of-00001.parquet"
#OUT = Path("./models/splits/train_val.parquet")
PKG = Path(__file__).resolve().parents[1]
REPO = Path(__file__).resolve().parents[3]
DATA = REPO / "data/talkpl-ai" # was Path("./data/talkpl-ai")
TRAIN_PARQUET = DATA / "TalkPlayData-Challenge-Dataset/data/train-00000-of-00001.parquet"
DEV_PARQUET = DATA / "TalkPlayData-Challenge-Dataset/data/test-00000-of-00001.parquet"
USER_EMB_TRAIN = DATA / "TalkPlayData-Challenge-User-Embeddings/data/train-00000-of-00001.parquet"
OUT = PKG / "models/splits/train_val.parquet" # was Path("./models/splits/..."
def hash_bucket(key: str, seed: int, mod: int = 100000) -> int:
"""Deterministic uniform hash of a string key into [0, mod)."""
h = hashlib.sha256(f"{seed}:{key}".encode()).hexdigest()
return int(h[:8], 16) % mod
def load_session_metadata(train_parquet: Path) -> pl.DataFrame:
"""Pull session_id, user_id, session_date, dominant (cat, spec) per session."""
print(f"Loading {train_parquet}")
full = pl.read_parquet(train_parquet)
rows = []
for r in full.to_dicts():
cg = r.get("conversation_goal") or {}
rows.append({
"session_id": r["session_id"],
"user_id": r["user_id"],
"session_date": str(r["session_date"]),
"category": cg.get("category") or "?",
"specificity": cg.get("specificity") or "?",
})
return pl.DataFrame(rows)
def load_warm_users(user_emb_train: Path) -> set[str]:
"""Users that have a non-empty cf-bpr embedding (i.e. real warm users)."""
if not user_emb_train.exists():
print(f" WARNING: {user_emb_train} missing — assuming all users warm.")
return set()
df = pl.read_parquet(user_emb_train)
warm = set()
for r in df.to_dicts():
v = r.get("cf-bpr")
if v is not None and len(v) > 0:
warm.add(r["user_id"])
return warm
def dominant_pair_per_user(sessions: pl.DataFrame) -> dict[str, tuple[str, str]]:
"""For each user, find the most common (category, specificity) pair.
Ties broken by alphabetical stability."""
cnt: dict[str, Counter] = defaultdict(Counter)
for r in sessions.iter_rows(named=True):
cnt[r["user_id"]][(r["category"], r["specificity"])] += 1
return {
uid: max(c.items(), key=lambda kv: (kv[1], kv[0]))[0]
for uid, c in cnt.items()
}
def sessions_per_user(sessions: pl.DataFrame) -> dict[str, int]:
return dict(zip(
sessions.group_by("user_id").len()["user_id"].to_list(),
sessions.group_by("user_id").len()["len"].to_list(),
))
def stratified_quota_sample(
candidates_by_stratum: dict[tuple, list[str]],
sessions_per_user_map: dict[str, int],
target_total_sessions: int,
seed: int,
) -> set[str]:
"""Pick users from each stratum proportional to its share of sessions in
`candidates_by_stratum`, until target_total_sessions is approximately reached.
Within each stratum, users are sorted by hash_bucket and taken in order so
the choice is deterministic for a given seed.
"""
strata_session_count = {
stratum: sum(sessions_per_user_map[u] for u in users)
for stratum, users in candidates_by_stratum.items()
}
total_pool_sessions = sum(strata_session_count.values())
if total_pool_sessions == 0:
return set()
chosen: set[str] = set()
chosen_sessions = 0
# Per-stratum target (round half-up)
for stratum, users in candidates_by_stratum.items():
share = strata_session_count[stratum] / total_pool_sessions
stratum_target = int(round(target_total_sessions * share))
if stratum_target == 0:
continue
ordered = sorted(users, key=lambda u: hash_bucket(u, seed + 7))
taken_sessions = 0
for u in ordered:
if taken_sessions >= stratum_target:
break
chosen.add(u)
taken_sessions += sessions_per_user_map[u]
chosen_sessions += sessions_per_user_map[u]
return chosen
def main():
p = argparse.ArgumentParser()
p.add_argument("--rerank-fraction", type=float, default=0.16)
p.add_argument("--val-fraction", type=float, default=0.20)
p.add_argument(
"--cutoff-date", type=str, default="2017-01-01",
help="Sessions on/after this date are eligible for train_rerank/val. "
"Softer than the v1 default (2018-01-01) so the holdout has a "
"richer distribution of session dates. Anything BEFORE the cutoff "
"stays in train_proj, which preserves temporal causality.",
)
p.add_argument("--cold-fraction", type=float, default=0.26,
help="Fraction of WARM holdout users to mark simulate_cold. "
"Real cold users are already cold; this dials the holdout's "
"effective cold rate to match dev's 25.8%.")
p.add_argument("--seed", type=int, default=42)
p.add_argument("--no-stratify-category", action="store_true",
help="Skip (category, specificity) stratification (not recommended).")
args = p.parse_args()
if args.rerank_fraction + args.val_fraction >= 1.0:
raise ValueError("rerank-fraction + val-fraction must be < 1.0")
# ── Load core tables ─────────────────────────────────────────────────────
sessions = load_session_metadata(TRAIN_PARQUET)
n_total = sessions.shape[0]
n_users = sessions["user_id"].n_unique()
print(f" {n_total} sessions, {n_users} users")
cutoff = args.cutoff_date
n_after = sessions.filter(pl.col("session_date") >= cutoff).shape[0]
print(f" cutoff={cutoff} post-cutoff sessions: {n_after} ({n_after/n_total:.3f})")
warm_users = load_warm_users(USER_EMB_TRAIN)
print(f" warm users (have cf-bpr): {len(warm_users)} / {n_users} "
f"({len(warm_users)/max(n_users,1):.3f})")
spu = sessions_per_user(sessions)
dom_pair = dominant_pair_per_user(sessions)
# ── Holdout eligibility: user's LATEST session must be on/after cutoff ──
user_latest = (
sessions
.group_by("user_id")
.agg(pl.col("session_date").max().alias("latest"))
)
eligible_users = set(
user_latest.filter(pl.col("latest") >= cutoff)["user_id"].to_list()
)
print(f" holdout-eligible users (latest session ≥ cutoff): {len(eligible_users)}")
# ── Stratified sampling of holdout users ─────────────────────────────────
target_holdout = int(n_total * (args.rerank_fraction + args.val_fraction))
print(f" target holdout sessions: {target_holdout}")
if args.no_stratify_category:
print(" category stratification: OFF")
candidates_by_stratum = {("ALL", "ALL"): list(eligible_users)}
else:
# One stratum per (warm, dominant_category, dominant_specificity) bucket.
# warm/cold included so we balance both axes simultaneously.
candidates_by_stratum: dict[tuple, list[str]] = defaultdict(list)
for u in eligible_users:
warm_flag = "W" if u in warm_users else "C"
cat, spec = dom_pair.get(u, ("?", "?"))
candidates_by_stratum[(warm_flag, cat, spec)].append(u)
print(f" category strata: {len(candidates_by_stratum)} buckets "
f"(warm/cold × cat × spec)")
holdout_users = stratified_quota_sample(
candidates_by_stratum, spu, target_holdout, args.seed,
)
holdout_sessions = sum(spu[u] for u in holdout_users)
print(f" selected holdout: {len(holdout_users)} users, {holdout_sessions} sessions "
f"({holdout_sessions/n_total:.3f} of total — target {target_holdout/n_total:.3f})")
# ── Within holdout, split rerank vs val proportionally per stratum ──────
rerank_share = args.rerank_fraction / (args.rerank_fraction + args.val_fraction)
rerank_threshold = int(rerank_share * 100000)
rerank_users = {
u for u in holdout_users
if hash_bucket(u, args.seed + 2) < rerank_threshold
}
val_users = holdout_users - rerank_users
rerank_sessions = sum(spu[u] for u in rerank_users)
val_sessions = sum(spu[u] for u in val_users)
print(f" rerank: {len(rerank_users)} users, {rerank_sessions} sessions "
f"({rerank_sessions/n_total:.3f})")
print(f" val: {len(val_users)} users, {val_sessions} sessions "
f"({val_sessions/n_total:.3f})")
# ── Cold simulation: only on WARM holdout users ─────────────────────────
warm_holdout = holdout_users & warm_users
cold_threshold = int(args.cold_fraction * 100000)
simulate_cold_users = {
u for u in warm_holdout
if hash_bucket(u, args.seed + 1) < cold_threshold
}
print(f" marked {len(simulate_cold_users)}/{len(warm_holdout)} WARM holdout "
f"users as simulate_cold (target {args.cold_fraction:.2f}, "
f"actual {len(simulate_cold_users)/max(len(warm_holdout),1):.3f})")
real_cold = len(holdout_users) - len(warm_holdout)
effective_cold = real_cold + len(simulate_cold_users)
print(f" effective cold rate in holdout: {effective_cold/max(len(holdout_users),1):.3f} "
f"(real {real_cold} + simulated {len(simulate_cold_users)})")
# ── Write output ─────────────────────────────────────────────────────────
rows = []
for r in sessions.iter_rows(named=True):
uid = r["user_id"]
if uid in rerank_users:
split = "train_rerank"
elif uid in val_users:
split = "val"
else:
split = "train_proj"
rows.append({
"session_id": r["session_id"],
"user_id": uid,
"split": split,
"simulate_cold": (split != "train_proj") and (uid in simulate_cold_users),
})
out_df = pl.DataFrame(rows)
# ── Diagnostics + final assertions ──────────────────────────────────────
print(f"\nFinal split counts:")
print(out_df.group_by("split").len().sort("split"))
proj_users = set(out_df.filter(pl.col("split") == "train_proj")["user_id"].to_list())
assert len(proj_users & rerank_users) == 0, "train_proj ∩ rerank not empty"
assert len(proj_users & val_users) == 0, "train_proj ∩ val not empty"
assert len(rerank_users & val_users) == 0, "rerank ∩ val not empty"
print("User-disjointness: OK")
# Temporal causality check
proj_sessions = sessions.filter(pl.col("user_id").is_in(proj_users))
rerank_sessions_df = sessions.filter(pl.col("user_id").is_in(rerank_users))
val_sessions_df = sessions.filter(pl.col("user_id").is_in(val_users))
if rerank_sessions_df.shape[0]:
max_proj = proj_sessions["session_date"].max()
min_rerank = rerank_sessions_df["session_date"].min()
min_val = val_sessions_df["session_date"].min() if val_sessions_df.shape[0] else None
print(f"\nTemporal causality:")
print(f" train_proj: {proj_sessions['session_date'].min()} .. {max_proj}")
print(f" rerank: {min_rerank} .. {rerank_sessions_df['session_date'].max()}")
if min_val:
print(f" val: {min_val} .. {val_sessions_df['session_date'].max()}")
# The cutoff we used guarantees: max(train_proj) ≤ cutoff ≤ min(rerank, val)
print(f" cutoff was {cutoff}; max(train_proj)={max_proj}, "
f"min(rerank)={min_rerank}, min(val)={min_val}")
if DEV_PARQUET.exists():
dev_dates = pl.read_parquet(DEV_PARQUET, columns=["session_date"])["session_date"].sort()
print(f" dev (reference): {dev_dates[0]} .. {dev_dates[-1]}")
OUT.parent.mkdir(parents=True, exist_ok=True)
out_df.write_parquet(OUT)
print(f"\nWrote {OUT}")
if __name__ == "__main__":
main() |