This project investigates the detection of AI-generated code versus human-written code, addressing academic integrity concerns in educational settings where generative models (e.g., ChatGPT, Codex) are increasingly used.

Traditional plagiarism tools like MOSS and TF-IDF clustering fail to distinguish between AI vs. human authorship. This notebook builds and compares several classification pipelines that utilize both lexical (TF-IDF) and structural (AST) features, integrating a contrastive learning network with a neural classifier for enhanced discrimination.

This notebook:

• Loads a labeled dataset of AI-generated vs. human-written code
• Extracts TF-IDF and AST-based vector features
• Trains machine learning classifiers (e.g., Random Forests)
• Implements a contrastive learning pipeline with triplet loss
• Evaluates models using standard classification metrics (Accuracy, F1, AUC)
• Visualizes learned embeddings using t-SNE and PCA

✅ Simply run the notebook top to bottom. All experiments and visualizations are included.

• data.jsonl: Raw dataset with labeled code snippets
  ➤ Download from GitHub

After downloading data.jsonl, place it in the notebook directory and run:

import json
import pandas as pd

with open('data.jsonl', 'r') as file:
    lines = file.readlines()

new_data = pd.DataFrame([json.loads(line) for line in lines])
display(new_data)

This will load the dataset into a DataFrame for further processing.

from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.decomposition import PCA
from sklearn.manifold import TSNE
from sklearn.model_selection import StratifiedKFold
from sklearn.metrics import (
    accuracy_score, precision_score, recall_score,
    f1_score, roc_curve, auc, classification_report, confusion_matrix
)
from sklearn.ensemble import RandomForestClassifier
import pandas as pd
import json

1. Open project_Cs6890_PLP.ipynb and run all cells.
2. The notebook will:
   • Train and validate models using stratified 2-fold cross-validation
   • Display classification metrics (Accuracy, Precision, Recall, F1-score, AUC)
   • Visualize embeddings using t-SNE and PCA

Models are evaluated using:

• Accuracy
• Precision
• Recall
• F1 Score
• AUC (Area Under the ROC Curve)

Developed for CS6890 - Programming Language Principles (PLP) at Utah State University