This project builds a moderation pipeline for Google-style reviews. Goal: automatically flag reviews that violate policy categories:

• advertisement: promotional content, referral links, sales pitches.
• irrelevant: unrelated to the business (random stories, off-topic text).
• no_visit_rant: rants/complaints where reviewer admits no visit.
• none: all other legitimate, on-topic reviews.

We compare multiple approaches:

1. Baseline (TF-IDF + Logistic Regression)
2. Zero-Shot Classification (Hugging Face NLI models)
3. Few-Shot Prompting (LLMs guided by src/prompts/policy.md)

The pipeline produces labeled datasets, model predictions, evaluation metrics, and prompt artifacts for analysis.

git clone cd review-moderation-ml

python3 -m venv .venv source .venv/bin/activate

pip install -U pip pip install pandas scikit-learn transformers torch accelerate datasets evaluate matplotlib rich joblib python-dotenv jupyter

Place raw Google Reviews dataset into data/raw/ For JSON data: python src/data/ingest_json.py

python src/models/hf_batch_infer.py

python src/models/baseline.py

python src/eval/visualise.py

Instructions and few-shot examples are in src/prompts/policy.md. Optionally, run a few-shot demo with a small instruction-tuned LLM.

• Sean See — Data ingestion, JSON/CSV cleaning pipeline, repository setup.
• Cyril Pedrina — Zero-shot classification pipeline (Hugging Face), pseudo-label generation.
• Jade Ng — Baseline model (TF-IDF + Logistic Regression), training, and evaluation metrics.
• Eleanor Cheak — Prompt engineering (policy.md), few-shot examples, and LLM experimentation.
• Zelda Seow — Documentation, and final presentation delivery.