Security fixes are applied to the latest revision on main. The project is an educational lab and does not currently publish long-term-supported release branches.

Do not disclose a vulnerability in a public issue before maintainers can assess it. Use GitHub's private vulnerability reporting for Doithoo/pytorch-text-classification-lab when available. Include affected revision, environment, reproduction steps, impact, and a minimal proof of concept without private data.

PyTorch .pt training checkpoints use pickle and may execute code while loading. The CLI deliberately describes .pt inputs as trusted. Never load an unknown user's training checkpoint or expose evaluate, export-inference, or resume as an unauthenticated upload service.

For inference distribution, use export-inference to create .safetensors weights and a JSON sidecar. This avoids executable pickle during weight loading, but provenance and integrity still matter. The sidecar contains the complete vocabulary and may be a derived data asset.

The download script verifies fixed SHA-256 values, but network and upstream repository trust still matter. Kaggle credentials and API tokens must stay outside the repository. Error-analysis files may contain complete source text; review privacy and dataset terms before sharing them.

This project is not a hardened serving system. It does not provide sandboxing, authentication, rate limiting, adversarial-input defenses, or calibrated safety decisions.