**Competition Rules**:
- **No external datasets or additional labeled data:** Participants must not use any external datasets (public, private, or proprietary) for training, validation, or model selection unless the instructor explicitly allows it in writing.
- **No external pretrained model weights:** Using pretrained model weights (transfer learning) that were trained on any external dataset is strictly prohibited. Models must be trained from scratch or using only the provided training data.
- **No external model checkpoints or code that would provide an unfair advantage:** Do not incorporate third-party model checkpoints, LLM/vision model outputs, or otherwise precomputed predictions into your training or inference pipeline.
- **No plagiarism:** All submitted code, reports, and notebooks must be the participant's own work. Copying, paraphrasing without attribution, or submitting work created by others (including online tutorials) is forbidden.
- **Required disclosures:** If you use any public libraries, code snippets, or algorithms (e.g., model architectures, loss functions), cite them clearly in your report or code comments. Failure to disclose sources may be treated as academic dishonesty.
- **Allowed tools:** Standard open-source libraries for building models (e.g., PyTorch, TensorFlow, scikit-learn, torchvision) and generic utilities (NumPy, Pandas, PIL/OpenCV) are permitted. Using pretrained weights, external datasets, or precomputed embeddings is not allowed unless explicitly permitted.

These rules are designed to ensure a fair and educational competition focused on model-building using only the provided data.