File size: 10,476 Bytes
f3101a4 |
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 |
"""
PubGuard β Multi-head Publication Gatekeeper
=============================================
Architecture
~~~~~~~~~~~~
βββββββββββββββ
β PDF text β
ββββββββ¬βββββββ
β
ββββββββΌβββββββ βββββββββββββββββββββ
β clean_text ββββββΊβ model2vec encode ββββΊ emb β R^512
βββββββββββββββ βββββββββββββββββββββ
β
βββββββββββββββββββΌββββββββββββββββββ
βΌ βΌ βΌ
βββββββββββββββββββ ββββββββββββββββ ββββββββββββββββ
β doc_type head β β ai_detect β β toxicity β
β (concat struct) β β head β β head β
β WΒ·[emb;feat]+b β β WΒ·emb + b β β WΒ·emb + b β
β β softmax(4) β β β softmax(2) β β β softmax(2) β
βββββββββββββββββββ ββββββββββββββββ ββββββββββββββββ
Each head is a single linear layer stored as a numpy .npz file
(weights W and bias b). Inference is pure numpy β no torch needed
at prediction time, matching the openalex classifier's deployment
philosophy.
The doc_type head additionally receives 14 structural features
(section headings present, citation density, etc.) concatenated
with the embedding β these are powerful priors that cost ~0 compute.
Performance target: β₯2,000 records/sec on CPU (same ballpark as
openalex classifier at ~3,000/sec).
"""
import logging
import time
from pathlib import Path
from typing import Any, Dict, List, Optional, Union
import numpy as np
from .config import PubGuardConfig, DOC_TYPE_LABELS, AI_DETECT_LABELS, TOXICITY_LABELS
from .text import clean_text, extract_structural_features, STRUCTURAL_FEATURE_NAMES, N_STRUCTURAL_FEATURES
logger = logging.getLogger(__name__)
class LinearHead:
"""
Single linear classifier head: logits = X @ W + b β softmax.
Stored as .npz with keys 'W', 'b', 'labels'.
"""
def __init__(self, labels: List[str]):
self.labels = labels
self.n_classes = len(labels)
self.W: Optional[np.ndarray] = None # (input_dim, n_classes)
self.b: Optional[np.ndarray] = None # (n_classes,)
def load(self, path: Path) -> bool:
if not path.exists():
return False
data = np.load(path, allow_pickle=True)
self.W = data["W"]
self.b = data["b"]
stored_labels = data.get("labels", None)
if stored_labels is not None:
self.labels = list(stored_labels)
self.n_classes = len(self.labels)
return True
def save(self, path: Path):
path.parent.mkdir(parents=True, exist_ok=True)
np.savez(path, W=self.W, b=self.b, labels=np.array(self.labels))
def predict(self, X: np.ndarray) -> tuple:
"""
Returns (pred_labels, pred_scores) for batch.
X : (batch, input_dim)
"""
logits = X @ self.W + self.b # (batch, n_classes)
probs = _softmax(logits) # (batch, n_classes)
pred_idx = np.argmax(probs, axis=1) # (batch,)
pred_scores = probs[np.arange(len(X)), pred_idx]
pred_labels = [self.labels[i] for i in pred_idx]
return pred_labels, pred_scores, probs
def _softmax(x: np.ndarray) -> np.ndarray:
"""Numerically stable softmax."""
e = np.exp(x - x.max(axis=-1, keepdims=True))
return e / e.sum(axis=-1, keepdims=True)
class PubGuard:
"""
Multi-head publication screening classifier.
Usage:
guard = PubGuard()
guard.initialize()
# Single document
verdict = guard.screen("Introduction: We present a novel ...")
# Batch
verdicts = guard.screen_batch(["text1", "text2", ...])
"""
def __init__(self, config: Optional[PubGuardConfig] = None):
self.config = config or PubGuardConfig()
self.model = None
self.head_doc_type = LinearHead(DOC_TYPE_LABELS)
self.head_ai_detect = LinearHead(AI_DETECT_LABELS)
self.head_toxicity = LinearHead(TOXICITY_LABELS)
self._initialized = False
# ββ Initialisation ββββββββββββββββββββββββββββββββββββββββββ
def initialize(self) -> bool:
"""Load embedding model + all classification heads."""
if self._initialized:
return True
logger.info("Initializing PubGuard...")
start = time.time()
self._load_model()
self._load_heads()
self._initialized = True
logger.info(f"PubGuard initialized in {time.time()-start:.1f}s")
return True
def _load_model(self):
"""Load model2vec StaticModel (same as openalex classifier)."""
from model2vec import StaticModel
cache = self.config.distilled_model_path
if cache.exists():
logger.info(f"Loading embedding model from {cache}")
self.model = StaticModel.from_pretrained(str(cache))
else:
logger.info(f"Downloading model: {self.config.model_name}")
self.model = StaticModel.from_pretrained(self.config.model_name)
cache.parent.mkdir(parents=True, exist_ok=True)
self.model.save_pretrained(str(cache))
logger.info(f"Cached to {cache}")
def _load_heads(self):
"""Load each classification head from .npz files."""
heads = [
("doc_type", self.head_doc_type, self.config.doc_type_head_path),
("ai_detect", self.head_ai_detect, self.config.ai_detect_head_path),
("toxicity", self.head_toxicity, self.config.toxicity_head_path),
]
for name, head, path in heads:
if head.load(path):
logger.info(f" Loaded {name} head: {path}")
else:
logger.warning(
f" {name} head not found at {path} β "
f"run `python -m pubguard.train` first"
)
# ββ Inference βββββββββββββββββββββββββββββββββββββββββββββββ
def screen(self, text: str) -> Dict[str, Any]:
"""Screen a single document. Returns verdict dict."""
return self.screen_batch([text])[0]
def screen_batch(self, texts: List[str]) -> List[Dict[str, Any]]:
"""
Screen a batch of documents.
Returns list of verdict dicts, each containing:
doc_type: {label, score}
ai_generated: {label, score}
toxicity: {label, score}
pass: bool (overall gate decision)
"""
if not self._initialized:
self.initialize()
if not texts:
return []
cfg = self.config
# ββ Preprocess ββββββββββββββββββββββββββββββββββββββββββ
cleaned = [clean_text(t, cfg.max_text_chars) for t in texts]
# ββ Embed βββββββββββββββββββββββββββββββββββββββββββββββ
embeddings = self.model.encode(cleaned)
norms = np.linalg.norm(embeddings, axis=1, keepdims=True)
norms = np.where(norms == 0, 1, norms) # avoid div-by-zero
embeddings = (embeddings / norms).astype("float32")
# ββ Structural features (for doc_type head) βββββββββββββ
struct_feats = np.array(
[list(extract_structural_features(t).values()) for t in cleaned],
dtype="float32",
)
doc_type_input = np.concatenate([embeddings, struct_feats], axis=1)
# ββ Per-head predictions ββββββββββββββββββββββββββββββββ
results = []
has_doc = self.head_doc_type.W is not None
has_ai = self.head_ai_detect.W is not None
has_tox = self.head_toxicity.W is not None
dt_labels, dt_scores, _ = (
self.head_doc_type.predict(doc_type_input) if has_doc
else (["unknown"] * len(texts), [0.0] * len(texts), None)
)
ai_labels, ai_scores, _ = (
self.head_ai_detect.predict(embeddings) if has_ai
else (["unknown"] * len(texts), [0.0] * len(texts), None)
)
tx_labels, tx_scores, _ = (
self.head_toxicity.predict(embeddings) if has_tox
else (["unknown"] * len(texts), [0.0] * len(texts), None)
)
for i in range(len(texts)):
# Gate logic
passes = True
if cfg.require_scientific and dt_labels[i] != "scientific_paper":
passes = False
if cfg.block_ai_generated and ai_labels[i] == "ai_generated":
passes = False
if cfg.block_toxic and tx_labels[i] == "toxic":
passes = False
results.append({
"doc_type": {
"label": dt_labels[i],
"score": round(float(dt_scores[i]), 4),
},
"ai_generated": {
"label": ai_labels[i],
"score": round(float(ai_scores[i]), 4),
},
"toxicity": {
"label": tx_labels[i],
"score": round(float(tx_scores[i]), 4),
},
"pass": passes,
})
return results
# ββ File-level convenience ββββββββββββββββββββββββββββββββββ
def screen_file(self, path: Path) -> Dict[str, Any]:
"""Read a text file and screen it."""
text = Path(path).read_text(errors="replace")
return self.screen(text)
|