File size: 9,888 Bytes
85de43a |
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 |
"""
Training pipeline for PubGuard classification heads.
Trains lightweight linear classifiers on frozen model2vec embeddings.
This follows the same paradigm as the openalex-topic-classifier:
the expensive embedding is pre-computed once, and the classifier
itself is a single matrix multiply β fast to train, fast to infer.
Training strategy:
1. Load + cache model2vec embeddings for all training data
2. For each head, fit a logistic regression (sklearn) with
class-balanced weights and L2 regularisation
3. Export weights as .npz for the numpy-only inference path
4. Report per-class precision / recall / F1 on held-out split
The entire pipeline trains in <5 minutes on CPU for ~50K samples,
consistent with your existing toolchain.
"""
import json
import logging
import time
from pathlib import Path
from typing import Dict, List, Optional, Tuple
import numpy as np
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import classification_report
from sklearn.model_selection import train_test_split
from .config import PubGuardConfig, DOC_TYPE_LABELS, AI_DETECT_LABELS, TOXICITY_LABELS
from .classifier import LinearHead
from .text import clean_text, extract_structural_features, N_STRUCTURAL_FEATURES
logger = logging.getLogger(__name__)
def load_ndjson(path: Path) -> Tuple[List[str], List[str]]:
"""Load NDJSON file β (texts, labels)."""
texts, labels = [], []
with open(path) as f:
for line in f:
if line.strip():
row = json.loads(line)
texts.append(row["text"])
labels.append(row["label"])
return texts, labels
def embed_texts(
texts: List[str],
config: PubGuardConfig,
cache_path: Optional[Path] = None,
) -> np.ndarray:
"""
Encode texts with model2vec, L2-normalise, return (N, D) float32.
Optionally caches to disk to avoid re-embedding on repeat runs.
"""
if cache_path and cache_path.exists():
logger.info(f"Loading cached embeddings from {cache_path}")
return np.load(cache_path)
from model2vec import StaticModel
model_path = config.distilled_model_path
if model_path.exists():
model = StaticModel.from_pretrained(str(model_path))
else:
model = StaticModel.from_pretrained(config.model_name)
model_path.parent.mkdir(parents=True, exist_ok=True)
model.save_pretrained(str(model_path))
logger.info(f"Embedding {len(texts)} texts...")
cleaned = [clean_text(t, config.max_text_chars) for t in texts]
embeddings = model.encode(cleaned, show_progress_bar=True)
# L2-normalise
norms = np.linalg.norm(embeddings, axis=1, keepdims=True)
norms = np.where(norms == 0, 1, norms)
embeddings = (embeddings / norms).astype("float32")
if cache_path:
cache_path.parent.mkdir(parents=True, exist_ok=True)
np.save(cache_path, embeddings)
logger.info(f"Cached embeddings to {cache_path}")
return embeddings
def compute_structural_features(texts: List[str]) -> np.ndarray:
"""Compute structural features for all texts."""
feats = []
for t in texts:
cleaned = clean_text(t)
feat_dict = extract_structural_features(cleaned)
feats.append(list(feat_dict.values()))
return np.array(feats, dtype="float32")
def train_head(
X_train: np.ndarray,
y_train: np.ndarray,
X_test: np.ndarray,
y_test: np.ndarray,
labels: List[str],
head_name: str,
C: float = 1.0,
max_iter: int = 1000,
) -> LinearHead:
"""
Train a single linear classification head.
Uses sklearn LogisticRegression with:
- L2 regularisation (C parameter)
- class_weight='balanced' for imbalanced data
- lbfgs solver (good for moderate feature counts)
- multinomial objective even for binary (consistent API)
Extracts W and b into a LinearHead for numpy-only inference.
"""
logger.info(f"\n{'='*60}")
logger.info(f"Training {head_name} head")
logger.info(f"{'='*60}")
logger.info(f" Train: {X_train.shape[0]:,} | Test: {X_test.shape[0]:,}")
logger.info(f" Features: {X_train.shape[1]} | Classes: {len(labels)}")
# Class distribution
unique, counts = np.unique(y_train, return_counts=True)
for u, c in zip(unique, counts):
logger.info(f" {u}: {c:,}")
start = time.time()
clf = LogisticRegression(
C=C,
max_iter=max_iter,
class_weight="balanced",
solver="lbfgs",
n_jobs=-1,
random_state=42,
)
clf.fit(X_train, y_train)
elapsed = time.time() - start
logger.info(f" Trained in {elapsed:.1f}s")
# Evaluate
y_pred = clf.predict(X_test)
report = classification_report(y_test, y_pred, target_names=labels, digits=4)
logger.info(f"\n{report}")
# Extract weights into LinearHead
head = LinearHead(labels)
# sklearn stores coef_ as (n_classes, n_features) for multinomial
# We want W as (n_features, n_classes) for X @ W + b
if clf.coef_.shape[0] == 1:
# Binary case: sklearn only stores one row
# Expand to full 2-class format
head.W = np.vstack([-clf.coef_[0], clf.coef_[0]]).T.astype("float32")
head.b = np.array([-clf.intercept_[0], clf.intercept_[0]], dtype="float32")
else:
head.W = clf.coef_.T.astype("float32") # (features, classes)
head.b = clf.intercept_.astype("float32")
# Sanity check: reproduce sklearn predictions
logits = X_test[:5] @ head.W + head.b
e = np.exp(logits - logits.max(axis=-1, keepdims=True))
probs = e / e.sum(axis=-1, keepdims=True)
np_pred_idx = np.argmax(probs, axis=1)
sk_pred_idx = clf.predict(X_test[:5]) # returns integer class indices
assert list(np_pred_idx) == list(int(x) for x in sk_pred_idx), \
f"Mismatch: {list(np_pred_idx)} vs {list(sk_pred_idx)}"
logger.info(" β Numpy inference matches sklearn predictions")
return head
def train_all(
data_dir: Path,
config: Optional[PubGuardConfig] = None,
test_size: float = 0.15,
):
"""
Train all three classification heads.
Args:
data_dir: Directory containing the prepared NDJSON files
config: PubGuard configuration
test_size: Fraction of data held out for evaluation
"""
config = config or PubGuardConfig()
data_dir = Path(data_dir)
cache_dir = data_dir / "embeddings_cache"
logger.info("=" * 60)
logger.info("PubGuard Training Pipeline")
logger.info("=" * 60)
logger.info(f"Data dir: {data_dir}")
logger.info(f"Models dir: {config.models_dir}")
start_total = time.time()
# ββ HEAD 1: doc_type ββββββββββββββββββββββββββββββββββββββββ
doc_type_path = data_dir / "doc_type_train.ndjson"
if doc_type_path.exists():
texts, labels = load_ndjson(doc_type_path)
label_to_idx = {l: i for i, l in enumerate(DOC_TYPE_LABELS)}
# Embed
embeddings = embed_texts(
texts, config,
cache_path=cache_dir / "doc_type_emb.npy",
)
# Add structural features
logger.info("Computing structural features...")
struct = compute_structural_features(texts)
X = np.concatenate([embeddings, struct], axis=1)
y = np.array([label_to_idx.get(l, 0) for l in labels])
X_tr, X_te, y_tr, y_te = train_test_split(
X, y, test_size=test_size, stratify=y, random_state=42
)
head = train_head(X_tr, y_tr, X_te, y_te, DOC_TYPE_LABELS, "doc_type")
head.save(config.doc_type_head_path)
logger.info(f"Saved β {config.doc_type_head_path}")
else:
logger.warning(f"doc_type data not found: {doc_type_path}")
# ββ HEAD 2: ai_detect βββββββββββββββββββββββββββββββββββββββ
ai_path = data_dir / "ai_detect_train.ndjson"
if ai_path.exists():
texts, labels = load_ndjson(ai_path)
label_to_idx = {l: i for i, l in enumerate(AI_DETECT_LABELS)}
embeddings = embed_texts(
texts, config,
cache_path=cache_dir / "ai_detect_emb.npy",
)
y = np.array([label_to_idx.get(l, 0) for l in labels])
X_tr, X_te, y_tr, y_te = train_test_split(
embeddings, y, test_size=test_size, stratify=y, random_state=42
)
head = train_head(X_tr, y_tr, X_te, y_te, AI_DETECT_LABELS, "ai_detect")
head.save(config.ai_detect_head_path)
logger.info(f"Saved β {config.ai_detect_head_path}")
else:
logger.warning(f"ai_detect data not found: {ai_path}")
# ββ HEAD 3: toxicity ββββββββββββββββββββββββββββββββββββββββ
tox_path = data_dir / "toxicity_train.ndjson"
if tox_path.exists():
texts, labels = load_ndjson(tox_path)
label_to_idx = {l: i for i, l in enumerate(TOXICITY_LABELS)}
embeddings = embed_texts(
texts, config,
cache_path=cache_dir / "toxicity_emb.npy",
)
y = np.array([label_to_idx.get(l, 0) for l in labels])
X_tr, X_te, y_tr, y_te = train_test_split(
embeddings, y, test_size=test_size, stratify=y, random_state=42
)
head = train_head(X_tr, y_tr, X_te, y_te, TOXICITY_LABELS, "toxicity")
head.save(config.toxicity_head_path)
logger.info(f"Saved β {config.toxicity_head_path}")
else:
logger.warning(f"toxicity data not found: {tox_path}")
elapsed = time.time() - start_total
logger.info(f"\nTotal training time: {elapsed/60:.1f} minutes")
logger.info("All heads saved to: " + str(config.models_dir))
|