File size: 7,757 Bytes
bd743a9 | 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 | """Train PIRD: a paraphrase-invariant, multi-signal detector.
Objective (the contribution):
- detection loss (human vs AI)
- augmentation: paraphrased AI is included, still labelled AI
- invariance: consistency loss pulling P(AI|x) ~ P(AI|paraphrase(x))
Full PIRD additionally fuses Stream A (statistical) + Stream C (stylometric) features onto the
encoder embedding, and fits a calibration temperature on a held-out val split (contribution C3).
Set use_features=False for the encoder-only "PIRD-lite" ablation.
"""
from __future__ import annotations
import json
import os
import random
import numpy as np
import torch
import torch.nn as nn
from torch.utils.data import Dataset, DataLoader
from transformers import AutoTokenizer
from .model import PIRDModel
from .attacks import Paraphraser
from .features import CombinedFeatures, standardize
class _PairDataset(Dataset):
def __init__(self, items, tok, max_len=256, use_features=False):
self.items = items; self.tok = tok; self.max_len = max_len; self.use_features = use_features
def __len__(self):
return len(self.items)
def __getitem__(self, i):
return self.items[i]
def collate(self, batch):
def enc(texts):
e = self.tok(texts, return_tensors="pt", truncation=True,
max_length=self.max_len, padding=True)
return e["input_ids"], e["attention_mask"]
ids, mask = enc([b["text"] for b in batch])
out = {"ids": ids, "mask": mask,
"labels": torch.tensor([b["label"] for b in batch], dtype=torch.float)}
if self.use_features:
out["extra"] = torch.tensor(np.stack([b["extra"] for b in batch]), dtype=torch.float)
has_para = [b for b in batch if b.get("para")]
if has_para:
p_ids, p_mask = enc([b["para"] for b in has_para])
out.update({"p_ids": p_ids, "p_mask": p_mask,
"p_labels": torch.tensor([b["label"] for b in has_para], dtype=torch.float)})
if self.use_features:
out["p_extra"] = torch.tensor(np.stack([b["p_extra"] for b in has_para]),
dtype=torch.float)
return out
def build_items(human, ai, recursive_rounds=0, seed=42, paraphraser=None):
para = paraphraser or Paraphraser()
print(f"paraphrasing {len(ai)} AI texts (rounds={max(1, recursive_rounds)}) ...")
ai_para = para.paraphrase_many(ai, rounds=max(1, recursive_rounds))
items = [{"text": h, "label": 0.0, "para": None} for h in human]
items += [{"text": a, "label": 1.0, "para": ap} for a, ap in zip(ai, ai_para)]
random.Random(seed).shuffle(items)
return items
def _attach_features(items, extractor, mean, std):
X = standardize(extractor.matrix([it["text"] for it in items]), mean, std)
for it, x in zip(items, X):
it["extra"] = x.astype("float32")
pidx = [i for i, it in enumerate(items) if it.get("para")]
if pidx:
P = standardize(extractor.matrix([items[i]["para"] for i in pidx]), mean, std)
for j, i in enumerate(pidx):
items[i]["p_extra"] = P[j].astype("float32")
def _fit_temperature(model, items, tok, max_len, device, use_features):
model.eval()
zs, ys = [], []
with torch.no_grad():
for i in range(0, len(items), 16):
chunk = items[i:i + 16]
e = tok([c["text"] for c in chunk], return_tensors="pt", truncation=True,
max_length=max_len, padding=True)
extra = (torch.tensor(np.stack([c["extra"] for c in chunk]), dtype=torch.float).to(device)
if use_features else None)
z = model(e["input_ids"].to(device), e["attention_mask"].to(device), extra)
zs.append(np.atleast_1d(z.cpu().numpy())); ys.append([c["label"] for c in chunk])
model.train()
z = np.concatenate(zs); y = np.concatenate(ys)
best_T, best = 1.0, 1e9
for T in np.linspace(0.5, 5.0, 91):
p = np.clip(1.0 / (1.0 + np.exp(-z / T)), 1e-6, 1 - 1e-6)
bce = -(y * np.log(p) + (1 - y) * np.log(1 - p)).mean()
if bce < best:
best, best_T = bce, float(T)
return best_T
def train_pird(items, out_dir="pird_ckpt", encoder="roberta-base", epochs=3, batch_size=8,
lr=2e-5, max_len=256, lam_inv=1.0, lam_aug=1.0, seed=42, device=None,
use_features=True, stat_model="gpt2", val_frac=0.15):
device = device or ("cuda" if torch.cuda.is_available() else "cpu")
torch.manual_seed(seed); np.random.seed(seed); random.seed(seed)
random.Random(seed).shuffle(items)
n_val = max(40, int(len(items) * val_frac))
val_items, train_items = items[:n_val], items[n_val:]
n_extra, mean, std = 0, None, None
if use_features:
extractor = CombinedFeatures(stat_model, device=device)
Xo = extractor.matrix([it["text"] for it in train_items])
mean, std = Xo.mean(0), Xo.std(0); std[std < 1e-6] = 1.0
n_extra = Xo.shape[1]
print(f"[pird] fusing {n_extra} A+C features (standardized)")
_attach_features(train_items, extractor, mean, std)
_attach_features(val_items, extractor, mean, std)
tok = AutoTokenizer.from_pretrained(encoder)
model = PIRDModel(encoder, n_extra=n_extra).to(device)
ds = _PairDataset(train_items, tok, max_len, use_features)
dl = DataLoader(ds, batch_size=batch_size, shuffle=True, collate_fn=ds.collate)
opt = torch.optim.AdamW(model.parameters(), lr=lr)
bce = nn.BCEWithLogitsLoss()
model.train()
for epoch in range(epochs):
run = {"det": 0.0, "aug": 0.0, "inv": 0.0}
for step, b in enumerate(dl):
extra = b.get("extra").to(device) if "extra" in b else None
logits = model(b["ids"].to(device), b["mask"].to(device), extra)
labels = b["labels"].to(device)
loss = bce(logits, labels)
run["det"] += loss.item()
if "p_ids" in b:
p_extra = b.get("p_extra").to(device) if "p_extra" in b else None
p_logits = model(b["p_ids"].to(device), b["p_mask"].to(device), p_extra)
p_labels = b["p_labels"].to(device)
aug = bce(p_logits, p_labels)
ai_logits = logits[labels == 1.0]
m = min(len(ai_logits), len(p_logits))
inv = ((torch.sigmoid(ai_logits[:m]) - torch.sigmoid(p_logits[:m])) ** 2).mean()
loss = loss + lam_aug * aug + lam_inv * inv
run["aug"] += aug.item(); run["inv"] += inv.item()
if not torch.isfinite(loss):
print(f" WARN non-finite loss at step {step}, skipping"); opt.zero_grad(); continue
opt.zero_grad(); loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
opt.step()
n = len(dl)
print(f"epoch {epoch+1}/{epochs} det={run['det']/n:.4f} aug={run['aug']/n:.4f} "
f"inv={run['inv']/n:.4f}")
temperature = _fit_temperature(model, val_items, tok, max_len, device, use_features)
print(f"[pird] calibration temperature T={temperature:.2f}")
os.makedirs(out_dir, exist_ok=True)
torch.save(model.state_dict(), os.path.join(out_dir, "pird.pt"))
cfg = {"encoder": encoder, "n_extra": n_extra, "max_len": max_len,
"use_features": use_features, "stat_model": stat_model, "temperature": temperature,
"feat_mean": (mean.tolist() if mean is not None else None),
"feat_std": (std.tolist() if std is not None else None)}
with open(os.path.join(out_dir, "config.json"), "w") as f:
json.dump(cfg, f)
print(f"saved PIRD checkpoint -> {out_dir}")
return out_dir
|