File size: 12,822 Bytes
17fc9a9 8505f8e 17fc9a9 320b589 17fc9a9 8505f8e 320b589 8505f8e 320b589 8505f8e 320b589 8505f8e 320b589 17fc9a9 8505f8e 320b589 8505f8e 17fc9a9 444e414 320b589 8505f8e 320b589 48db85f 320b589 444e414 320b589 444e414 320b589 8505f8e 17fc9a9 8505f8e 320b589 8505f8e 17fc9a9 8505f8e | 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 | """Stage 3: mint (direction, target_texts) pairs — with REAL, VALIDATED probes.
For each sampled cluster pair (A,B) we TRAIN a logistic-regression probe on a train split of the
READ_LAYER residuals and only keep the direction if it actually works:
- held-out separation AUC(A vs B) >= --min-auc (else the "concept direction" is meaningless)
- cluster-A val residuals project HIGH on the unit probe and clear cluster-B by a margin
(else the direction isn't something the corpus strongly exhibits → nothing to maximize).
The kept unit direction is the conditioning vector; targets = centroid-closest cluster-A texts.
We log the AUC / projection-margin distributions and the fraction of pairs dropped.
OUT-OF-CORE: residuals come straight from the clustering stage's emb.f32 ([n_docs, d] memmap,
row-aligned with assign.npy / texts.jsonl — no separate resid cache). Cluster member rows are
gathered on demand through a bounded LRU (--cache-gb), texts through a byte-offset index, so RAM
never scales with corpus size. Probes are GPU-batched logistic regressions: --probe-batch pairs
fit in parallel per exact Newton/IRLS run (sklearn's objective: sum-BCE + 0.5/C·||w||²) —
replaces sklearn-per-pair, which is far too slow at millions of pairs.
python scripts/build_data.py --acts-dir data/actclusters --n-examples 8000000 \
--min-auc 0.9 --min-margin 1.0 --hard-neg-k 16
"""
import argparse
import json
import os
import numpy as np
import torch
from mxf.config import D_MODEL, BuildDataConfig
class TextBank:
"""Random access into texts.jsonl by row via a byte-offset index (corpus never in RAM)."""
def __init__(self, path):
self.f = open(path, "rb")
off = [0]
for line in self.f:
off.append(off[-1] + len(line))
self.off = np.asarray(off[:-1], dtype=np.int64)
def __getitem__(self, i):
self.f.seek(self.off[i])
return json.loads(self.f.readline())["t"]
class ResidLRU:
"""Bounded LRU over per-cluster member-row gathers from the emb memmap."""
def __init__(self, emb, budget_gb):
self.emb, self.d, self.used, self.max = emb, {}, 0, int(budget_gb * 1e9)
def get(self, c, rows):
v = self.d.pop(c, None)
if v is None:
v = np.asarray(self.emb[rows])
self.used += v.nbytes
while self.used > self.max and self.d:
self.used -= self.d.pop(next(iter(self.d))).nbytes
self.d[c] = v # (re)insert = most recent
return v
@torch.no_grad()
def fit_probes_gpu(feats, C, iters, device):
"""One L2 logistic regression per (Atr, Btr, Ava, Bva) tuple, ALL fitted in parallel on GPU.
Exact Newton/IRLS in the n-dim dual: with n<=~128 samples << d=4096, w* = Xᵀα (representer),
and Woodbury keeps every solve at [P, n, n] fp64 — matches sklearn's optimum of
sum-BCE + 0.5/C·||w||² (+ ~free intercept, δ=1e-4) in ~10 iterations. Padded batch + masks;
val AUC / medians batched too. Returns per pair (unit_w, val_auc, projA_med, projB_med) or
None (degenerate)."""
P, d = len(feats), feats[0][0].shape[1]
nt = max(len(A) + len(B) for A, B, _, _ in feats)
va = max(1, max(len(v) for _, _, v, _ in feats)); vb = max(1, max(len(v) for _, _, _, v in feats))
X = np.zeros((P, nt, d), np.float32); y = np.zeros((P, nt), np.float64); m = np.zeros((P, nt), np.float64)
Xa = np.zeros((P, va, d), np.float32); ma = np.zeros((P, va), bool)
Xb = np.zeros((P, vb, d), np.float32); mb = np.zeros((P, vb), bool)
for i, (A, B, Av, Bv) in enumerate(feats): # pad on CPU, one H2D per tensor (not per pair)
X[i, : len(A)] = A; X[i, len(A) : len(A) + len(B)] = B
y[i, : len(A)] = 1.0; m[i, : len(A) + len(B)] = 1.0
Xa[i, : len(Av)] = Av; ma[i, : len(Av)] = True
Xb[i, : len(Bv)] = Bv; mb[i, : len(Bv)] = True
X, y, m, Xa, ma, Xb, mb = (torch.from_numpy(t).to(device) for t in (X, y, m, Xa, ma, Xb, mb))
G = (C * (X @ X.transpose(1, 2))).double() + 1e4 # X̃D⁻¹X̃ᵀ with X̃=[X,1], D=diag(1/C..,δ)
eye = torch.eye(nt, device=device, dtype=torch.float64)
z = torch.zeros(P, nt, device=device, dtype=torch.float64)
c = torch.zeros_like(z)
for _ in range(iters): # damped-free Newton: θ' = -D⁻¹X̃ᵀc
p = torch.sigmoid(z)
r = (p - y) * m
sh = ((p * (1 - p)).clamp(min=1e-12) * m).sqrt()
q = (G @ r.unsqueeze(-1)).squeeze(-1) + z # X̃D⁻¹∇: pads contribute 0 (r,c=0 there)
A_ = eye + sh.unsqueeze(-1) * G * sh.unsqueeze(-2)
v = torch.linalg.solve(A_, (sh * q).unsqueeze(-1)).squeeze(-1)
c = r - sh * v
z = -(G @ c.unsqueeze(-1)).squeeze(-1)
w = -C * torch.einsum("pn,pnd->pd", c.float(), X)
nrm = w.norm(dim=1)
wu = w / nrm[:, None].clamp(min=1e-12)
pa = torch.einsum("pvd,pd->pv", Xa, wu).masked_fill(~ma, torch.nan)
pb = torch.einsum("pvd,pd->pv", Xb, wu).masked_fill(~mb, torch.nan)
both = ma[:, :, None] & mb[:, None, :]
gt = (pa[:, :, None] > pb[:, None, :]).double() + 0.5 * (pa[:, :, None] == pb[:, None, :]).double()
auc = (gt.where(both, 0.0)).sum((1, 2)) / both.sum((1, 2)).clamp(min=1)
medA, medB = pa.nanquantile(0.5, dim=1), pb.nanquantile(0.5, dim=1)
nrm, WU, auc, medA, medB = (t.cpu().numpy() for t in (nrm, wu, auc, medA, medB))
return [None if (nrm[i] < 1e-8 or not len(f[2]) or not len(f[3]))
else (WU[i], float(auc[i]), float(medA[i]), float(medB[i])) for i, f in enumerate(feats)]
def main():
cfg = BuildDataConfig()
ap = argparse.ArgumentParser()
ap.add_argument("--acts-dir", default="data/actclusters",
help="embed_cluster_acts out-dir (emb.f32 / assign.npy / centroids.npy / texts.jsonl)")
ap.add_argument("--out-dir", default=cfg.out_dir)
ap.add_argument("--n-examples", type=int, default=cfg.n_examples)
ap.add_argument("--targets", type=int, default=cfg.targets_per_example)
ap.add_argument("--min-auc", type=float, default=0.9, help="drop pairs the probe can't separate")
ap.add_argument("--min-margin", type=float, default=1.0,
help="min (projA_med - projB_med) in resid units: cluster must activate the probe")
ap.add_argument("--val-frac", type=float, default=0.3)
ap.add_argument("--probe-c", type=float, default=cfg.probe_c)
ap.add_argument("--hard-neg-k", type=int, default=0,
help="0 = random B (trivially-separable, cartoon directions). >0 = draw B from "
"A's k nearest clusters by centroid → subtle, information-rich probe directions.")
ap.add_argument("--members-cap", type=int, default=64,
help="residual rows per cluster used to fit/val the probe")
ap.add_argument("--probe-batch", type=int, default=512, help="cluster pairs fitted per GPU LR batch")
ap.add_argument("--probe-iters", type=int, default=10, help="Newton/IRLS iterations per LR fit")
ap.add_argument("--cache-gb", type=float, default=32.0, help="LRU budget for member-residual gathers")
ap.add_argument("--device", default="cuda")
a = ap.parse_args()
os.makedirs(a.out_dir, exist_ok=True)
rng = np.random.default_rng(cfg.seed)
meta = json.load(open(f"{a.acts_dir}/meta.json"))
K = meta["clusters"]
emb = np.memmap(f"{a.acts_dir}/emb.f32", dtype=np.float32, mode="r",
shape=(meta["n_docs"], meta["d"])) # -1 is illegal in memmap shapes
assign = np.load(f"{a.acts_dir}/assign.npy")
cent = np.load(f"{a.acts_dir}/centroids.npy")
texts = TextBank(f"{a.acts_dir}/texts.jsonl")
# group doc rows by cluster CSR-style: rows(c) = order[startx[c]:startx[c+1]] (no 1M py lists)
order = np.argsort(assign, kind="stable")
startx = np.zeros(K + 1, dtype=np.int64)
np.cumsum(np.bincount(assign, minlength=K), out=startx[1:])
# per usable cluster (>=8 members): shuffled member rows (probe fit/val, <= members-cap),
# train/val split, target rows = centroid-closest among docs NOT in the probe-fit split —
# targets come only from docs the probe never saw, so direction↔text can't be memorized
mrows, split, tgt_rows, mus = {}, {}, {}, {}
for c in range(K):
rows = order[startx[c] : startx[c + 1]]
if len(rows) < 8:
continue
perm = rows[rng.permutation(len(rows))]
mem = perm[: a.members_cap]
s = max(2, int(len(mem) * (1 - a.val_frac)))
cand = np.sort(perm[s:]) # all cluster docs minus probe-fit rows
if not len(cand):
continue
dctr = np.linalg.norm(np.asarray(emb[cand]) - cent[c], axis=1)
mrows[c], split[c], tgt_rows[c] = mem, s, cand[np.argsort(dctr)[:32]].tolist()
if a.hard_neg_k > 0:
mus[c] = np.asarray(emb[np.sort(mem[:s])]).mean(0)
if len(mrows) % 50_000 == 0:
print(f" prepped {len(mrows)} clusters", flush=True)
pool = np.array(sorted(mrows))
nid = {int(c): i for i, c in enumerate(pool)}
print(f"{len(pool)} clusters with >=8 members", flush=True)
# hard-negative pairing: nearest clusters in QWEN3 layer-27 residual space (train-split mean
# resid) — the probe/injection/reward live in Qwen3 residuals, so the genuinely hard negatives
# are pairs the model itself struggles to linearly separate. GPU brute-force top-k.
nearest = None
if a.hard_neg_k > 0:
mu = torch.from_numpy(np.stack([mus[c] for c in pool])).to(a.device)
mn, mub = (mu * mu).sum(1), mu.to(torch.bfloat16)
nearest = np.empty((len(pool), min(a.hard_neg_k, len(pool) - 1)), dtype=np.int64)
cb = max(64, int(4e9 / (len(pool) * 4)))
for s0 in range(0, len(pool), cb):
d2 = mn[None] - 2 * (mub[s0 : s0 + cb] @ mub.T).float()
d2[torch.arange(len(d2)), torch.arange(s0, s0 + len(d2))] = torch.inf # exclude self
nearest[s0 : s0 + len(d2)] = d2.topk(nearest.shape[1], dim=1, largest=False).indices.cpu().numpy()
nearest = pool[nearest] # pool index -> cluster id
del mu, mub, mn, d2
torch.cuda.empty_cache()
lru = ResidLRU(emb, a.cache_gb)
vec_bank = np.memmap(f"{a.out_dir}/vecs.f32", dtype=np.float32, mode="w+",
shape=(a.n_examples, D_MODEL))
recs = open(f"{a.out_dir}/records.jsonl", "w")
aucs, margins = [], []
n, tried, dropped = 0, 0, 0
while n < a.n_examples:
pairs, feats = [], []
while len(pairs) < a.probe_batch: # sample a GPU batch of cluster pairs
A = int(rng.choice(pool))
B = int(rng.choice(nearest[nid[A]])) if a.hard_neg_k > 0 else int(rng.choice(pool))
if B == A:
continue
RA, RB = lru.get(A, mrows[A]), lru.get(B, mrows[B])
sA, sB = split[A], split[B]
pairs.append((A, B))
feats.append((RA[:sA], RB[:sB], RA[sA:], RB[sB:]))
for (A, B), pr in zip(pairs, fit_probes_gpu(feats, a.probe_c, a.probe_iters, a.device)):
tried += 1
if tried % 25_000 == 0:
med = (f"auc med {np.median(aucs):.3f} margin med {np.median(margins):.2f}"
if aucs else "none kept yet")
print(f"minted {n}/{a.n_examples} | kept {tried-dropped}/{tried} pairs "
f"(drop {dropped/tried:.0%}) | {med}", flush=True)
if pr is None:
dropped += 1; continue
wu, auc, pA, pB = pr
if auc < a.min_auc or (pA - pB) < a.min_margin: # weak probe OR cluster doesn't activate it
dropped += 1; continue
aucs.append(auc); margins.append(pA - pB)
for tr in rng.permutation(tgt_rows[A])[: a.targets]:
if n >= a.n_examples:
break
vec_bank[n] = wu
recs.write(json.dumps({"vec_idx": n, "target_text": texts[int(tr)][:1200],
"cluster": A, "val_auc": round(auc, 3),
"proj_margin": round(pA - pB, 2)}) + "\n")
n += 1
if n >= a.n_examples:
break
recs.close(); vec_bank.flush()
stats = {"n_examples": n, "pairs_tried": tried, "pairs_dropped": dropped,
"drop_frac": dropped / max(tried, 1), "auc_median": float(np.median(aucs)),
"auc_p10": float(np.percentile(aucs, 10)), "margin_median": float(np.median(margins)),
"min_auc": a.min_auc, "min_margin": a.min_margin}
json.dump(stats, open(f"{a.out_dir}/build_stats.json", "w"), indent=1)
print(f"BUILD_DATA_DONE {stats}", flush=True)
if __name__ == "__main__":
main()
|