Laziestant • AI Text Detection API

Flask API + web UI to classify text as AI‑generated or human‑written using TF‑IDF and Logistic Regression. Includes CSV upload, training, prediction, and a model status panel with charts.

What it does

• Upload a CSV dataset and train a text classifier
• Predict per‑text probabilities (AI vs Human) via web UI or JSON
• Persist model artifacts and metadata to models/
• Visualize model accuracy (pie chart) and data split (radar chart) in the UI

Tech stack

• Backend: Flask, scikit‑learn
• Model: LogisticRegression (liblinear, class_weight="balanced")
• Vectorizer: TfidfVectorizer (norm='l1', 1–2 grams, sublinear_tf, stop_words='english')
• Preprocessing: lowercasing, newline/quote cleanup, contraction expansion, punctuation removal, whitespace collapse
• Frontend: Tailwind CSS, Chart.js

Endpoints

• GET / – Web UI (analyzer + model status)
• GET /api/info – Project + model info and metadata
• GET /model/status – Detailed model status (features, file sizes)
• GET|POST /upload-csv – Upload CSV dataset
• POST /train – Train on the latest uploaded CSV
• POST /predict – Predict label and probabilities for input text

Quickstart

1. Install dependencies

python -m venv .venv && source .venv/Scripts/activate
pip install -r requirements.txt

2. Run the app

py app.py
# then open http://localhost:5000

Optional: Windows run.bat or Unix run.sh are included.

Dataset format

• Preferred: CSV with columns text, label
• Supported labels: human/ai, or 0/1 (common synonyms are normalized)
• Headerless CSVs are supported: last column is the label, all prior columns are joined as text

Example rows:

text,label
This essay discusses...,human
AI systems often...,ai

Train and predict

After uploading a CSV (via UI or /upload-csv), trigger a training run:

curl -X POST http://127.0.0.1:5000/train

Make a prediction:

curl -X POST http://127.0.0.1:5000/predict \
	-H "Content-Type: application/json" \
	-d '{"text":"Write a short paragraph about the solar system."}'

Response (abridged):

{
	"prediction": "human",
	"confidence": 0.51,
	"probabilities": { "human": 0.51, "ai": 0.49 }
}

UI and charts

• Model Status shows: Status, Algorithm, Health (High/Medium/Low), Accuracy, Dataset Size, Last Trained
• Charts:
  • Accuracy: pie chart (Correct vs Other)
  • Data Split: radar chart (Total / Train / Test)

Artifacts and metadata

Saved under models/:

• model.pkl – trained Logistic Regression
• vectorizer.pkl – TF‑IDF vectorizer
• label_mapping.pkl – normalized label→class mapping
• model_meta.json – metadata: accuracy, dataset size, split counts, vectorizer config, model params, dataset_used
• latest_dataset.txt – path to the most recent dataset

Uploads are stored under uploads/.

Configuration

• MAX_CONTENT_MB (env): max upload size in MB (default 100)

Troubleshooting

• 400 “Dataset must have exactly 2 labels”: ensure only two classes exist in label
• 413 “File too large”: reduce file size or increase MAX_CONTENT_MB
• Headerless CSVs: ensure the last column is the label; others are text parts

Notes

• The UI “confidence” is per‑text; the Model Status “accuracy” is overall test accuracy.