IMDb情感分析全流程：DistilBERT LoRA与TF-IDF基线对比

In this tutorial, we develop an end-to-end sentiment analysis workflow using the Stanford NLP IMDb Large Movie Review Dataset and compare classical machine learning with parameter-efficient transformer fine-tuning. We begin by establishing a reproducible environment and auditing the dataset for class ordering, review-length skew, duplicate leakage, and preprocessing artifacts before training a strong TF-IDF and Logistic Regression baseline. We then fine-tune DistilBERT with LoRA through PEFT, evaluate it using accuracy, macro-F1, ROC-AUC, confusion matrices, and ROC curves, and examine threshold selection and probability calibration through Expected Calibration Error and reliability analysis. Beyond headline metrics, we investigate confident errors, performance across review lengths, word-level occlusion saliency, and head-versus-tail truncation to understand how the model reaches its predictions and where long-context limitations affect performance. Finally, we use the unlabeled IMDb split for confidence-based pseudo-labeling, compare the resulting semi-supervised model against our baseline, and save the merged transformer for reusable sentiment inference.

import importlib.util, subprocess, sys, os, time, random, warnings, inspect, hashlib
warnings.filterwarnings("ignore")
os.environ["TOKENIZERS_PARALLELISM"] = "false"
os.environ["WANDB_DISABLED"] = "true"
_REQUIRED = {
   "transformers": "transformers",
   "datasets": "datasets",
   "peft": "peft",
   "accelerate": "accelerate",
   "sklearn": "scikit-learn",
}
_missing = [pkg for mod, pkg in _REQUIRED.items() if importlib.util.find_spec(mod) is None]
if _missing:
   print(f"Installing: {', '.join(_missing)} ...")
   subprocess.run([sys.executable, "-m", "pip", "install", "-q", *_missing], check=True)
   print("Done. (If imports fail below, restart the runtime and re-run.)\n")
import numpy as np
import pandas as pd
import torch
import matplotlib.pyplot as plt
from datasets import load_dataset
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.linear_model import LogisticRegression
from sklearn.pipeline import make_pipeline
from sklearn.metrics import (accuracy_score, f1_score, roc_auc_score,
                            classification_report, confusion_matrix, roc_curve)
from transformers import (AutoTokenizer, AutoModelForSequenceClassification,
                         TrainingArguments, Trainer, DataCollatorWithPadding,
                         EarlyStoppingCallback, set_seed)
from peft import LoraConfig, get_peft_model, TaskType
def _disable_torchao_probe():
   patched = []
   try:
       import peft.import_utils as _piu
       _piu.is_torchao_available = lambda: False
       patched.append("peft.import_utils")
   except Exception:
       pass
   for _name, _mod in list(sys.modules.items()):
       if _name.startswith("peft") and hasattr(_mod, "is_torchao_available"):
           _mod.is_torchao_available = lambda: False
           patched.append(_name)
   return patched
try:
   import torchao as _tao
   _v = getattr(_tao, "__version__", "?")
   if tuple(int(x) for x in _v.split(".")[:2]) < (0, 16):
       print(f"[compat] torchao {_v} < 0.16 -> disabling PEFT's torchao probe: "
             f"{', '.join(_disable_torchao_probe())}")
except Exception:
   _disable_torchao_probe()
SEED        = 42
MODEL_NAME  = "distilbert-base-uncased"
MAX_LEN     = 256
N_TRAIN     = 5000
N_EVAL      = 2000
N_UNSUP     = 3000
EPOCHS      = 2
BATCH       = 16
LR          = 3e-4
FULL_RUN    = False
if FULL_RUN:
   N_TRAIN, N_EVAL, EPOCHS = 25000, 25000, 3
set_seed(SEED); random.seed(SEED); np.random.seed(SEED); torch.manual_seed(SEED)
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
print("=" * 79)
print(f"device={DEVICE} | torch={torch.__version__} | "
     f"gpu={torch.cuda.get_device_name(0) if DEVICE=='cuda' else 'n/a'}")
print("=" * 79)
t0 = time.time()
raw = load_dataset("stanfordnlp/imdb")
print(raw, f"\nloaded in {time.time()-t0:.1f}s\n")
print("--- example (truncated) ---")
print("label:", raw["train"][0]["label"], "|", raw["train"][0]["text"][:300], "...\n")
first_labels = np.array(raw["train"]["label"][:5])
last_labels  = np.array(raw["train"]["label"][-5:])
print(f"TRAP #1 - split ordering: first 5 labels {first_labels}, "
     f"last 5 labels {last_labels}  -> ALWAYS shuffle before subsampling.")
train_full = raw["train"].shuffle(seed=SEED)
test_full  = raw["test"].shuffle(seed=SEED)
train_ds   = train_full.select(range(min(N_TRAIN, len(train_full))))
eval_ds    = test_full.select(range(min(N_EVAL, len(test_full))))
print(f"   after shuffle+subsample: train balance = "
     f"{np.bincount(train_ds['label'])}, eval balance = {np.bincount(eval_ds['label'])}")
lens = np.array([len(t.split()) for t in train_full["text"]])
q = np.percentile(lens, [50, 75, 90, 95, 99])
print(f"\nTRAP #2 - length (words): median={q[0]:.0f} p75={q[1]:.0f} p90={q[2]:.0f} "
     f"p95={q[3]:.0f} p99={q[4]:.0f} max={lens.max()}")
print(f"   ~{(lens > MAX_LEN*0.75).mean()*100:.1f}% of reviews exceed MAX_LEN={MAX_LEN} "
     f"tokens (rough words->tokens factor 1.3). Section 9 measures what that costs.")
h_tr = {hashlib.md5(t.encode()).hexdigest() for t in raw["train"]["text"]}
h_te = {hashlib.md5(t.encode()).hexdigest() for t in raw["test"]["text"]}
print(f"\nTRAP #3 - leakage: {len(h_tr & h_te)} exact duplicate reviews across "
     f"train/test; {len(raw['train'])-len(h_tr)} dupes inside train itself.")
def clean(t):
   return t.replace("<br />", " ").replace("<br/>", " ").strip()
plt.figure(figsize=(11, 3.2))
plt.subplot(1, 2, 1)
plt.hist(np.clip(lens, 0, 1000), bins=60)
plt.axvline(MAX_LEN, ls="--", color="k", label=f"MAX_LEN={MAX_LEN}")
plt.title("Review length (words, clipped at 1000)"); plt.legend()
plt.subplot(1, 2, 2)
plt.bar(["neg", "pos"], np.bincount(raw["train"]["label"]))
plt.title("Train class balance (perfectly balanced)")
plt.tight_layout(); plt.show()

We configure the Colab environment, install the required libraries, apply the PEFT–torchao compatibility fix, and set deterministic seeds for reproducible experiments. We load the Stanford IMDb dataset, shuffle and subsample the train and test splits, and inspect class balance, review-length distributions, duplicate leakage, and HTML artifacts. We also visualize review lengths and label frequencies so we understand the dataset structure before building any models.