CIL 2026 — Sentiment Classification

Bilingual (DE/EN) product-review star-rating prediction.

Setup (macOS)

1. Prerequisites

• macOS with Python 3.9+ (python3 --version to check)
• Git
• An nethz account with cluster access (to grab the data)

2. Clone

git clone <repo-url> cil
cd cil

3. Create the virtual environment

python3 -m venv .venv
source .venv/bin/activate
pip install --upgrade pip
pip install -r requirements.txt

Activate the env every new terminal session: source .venv/bin/activate.

4. Get the data

The CSVs are not committed (too large). Copy them from the cluster:

mkdir -p data
scp <your-nethz>@student-cluster.inf.ethz.ch:/cluster/courses/cil/text-classification/data/train.csv data/
scp <your-nethz>@student-cluster.inf.ethz.ch:/cluster/courses/cil/text-classification/data/test.csv  data/

You should end up with data/train.csv (~58 MB) and data/test.csv (~38 MB).

5. Verify

python -c "
import pandas as pd
print(pd.read_csv('data/train.csv').shape)
"

Expect: (252000, 3).

Project layout

code/
├── data/                          # CSVs + cached features (gitignored)
├── submissions/                   # Kaggle CSVs
├── eda.ipynb                      # exploratory data analysis
├── b1_tfidf_logreg.ipynb          # baseline 1: TF-IDF + LogReg
├── b2_sentemb_logreg.ipynb        # baseline 2: frozen sentence-transformer + LogReg
├── requirements.txt
└── README.md

Running notebooks

In VS Code: open the .ipynb, click Select Kernel → pick the one at .venv/bin/python.

In Jupyter: jupyter notebook <file>.ipynb.

Always run b1_tfidf_logreg.ipynb first — it creates data/val_indices.npy, the canonical 90/10 split that every later notebook reuses.

Notes

• The first run of b2_sentemb_logreg.ipynb downloads ~470 MB of model weights and encodes all 420K reviews. Encoding is cached to data/b2_*.npy so subsequent runs are instant.
• On Apple Silicon (M1+) the encoder uses MPS automatically. On Intel Macs it falls back to CPU (~2 h for the encode).
• The cluster is for the main transformer fine-tune, not the baselines. Separate doc.