1) Overview Purpose. Skills-based competition to build a CatBoost-only model that identifies a subset of tokens from a larger dataset. Agreement. By registering and submitting (“Submission”), you agree to these Official Rules (“Rules”) and any instructions posted on the Competition Website. Sponsor. [Sponsor] (“Sponsor”). Competition Website & Contact. https://discord.gg/cabApvps 2) Prizes (Total: 6000 USDC) 1st: 3,000 USDC 2nd: 1,500 USDC 3rd: 1,000 USDC 4th: 500 USDC Prizes paid in USDC to a wallet provided by the winner(s). Any taxes/compliance requirements are the responsibility of winners. 3) Key Requirements (CatBoost Only) Only CatBoost may be used for learning (training and inference). Allowed: feature engineering; preprocessing; CatBoost params; CV; class weights; early stopping; ensembling only of CatBoost models. Not allowed: any non-CatBoost learners (e.g., XGBoost, LightGBM, neural nets, linear models), meta-models outside CatBoost, manual hand-labeling or heuristics leveraging information not permitted by these Rules. 4) Data Use Competition Data only. No additional/external data is permitted (public, private, or otherwise). No redistribution. Do not share Competition Data outside the contest. 5) Teams & Accounts Team size: up to 5 people. Team changes/mergers: allowed until Sponsor announces a team-freeze in Discord (or, if none is announced, before you make any Final Submission). One account per person. No multiple accounts or participation on multiple teams. 6) Submissions Daily limit: up to 5 Submissions per team per calendar day. Final for judging: you may designate up to 2 Submissions as Final before the deadline announced in Discord (if no explicit deadline is posted, the contest ends when Sponsor closes submissions). Only Final Submissions are eligible for prizes. 7) Packaging (reproducibility required): Training + inference code, deterministic seeds, environment/requirements, run script, and instructions to reproduce the Final Submission. If artifacts are needed, include them or provide deterministic steps to rebuild. A brief Model Card/README describing features, preprocessing, CatBoost version/params, validation scheme, and expected runtime. You must keep a Notebook (or repo) that can recreate your scored submission file end-to-end. Final prize eligibility requires code review. Within 72 hours of request, finalists must provide: The exact notebook(s)/scripts used to generate the submitted CSV Environment details (Python version, key libraries + versions) Determinism: seeds set for all randomness, or an explanation why results may vary Data sources: all external data must be publicly available and cited (with links), and permitted by Kaggle rules Inference instructions: how to run from raw competition data to final submission Runtime: expected run time and any hardware needs (GPU/TPU/CPU) How to share code for review (pick one): Make your Kaggle Notebook Public and link it in your submission description, or Keep it Private and share with hosts (Share → add organizer usernames), or Provide a downloadable snapshot (e.g., .ipynb + requirements.txt) to the organizers. Disqualification: We reserve the right to disqualify entries we cannot reproduce, that use prohibited data, or that fail to provide code on request. Winners: must provide the exact code and instructions that reproduce the winning score on the public/private tests. 8) Task & Evaluation Task. Predict the ground-truth subset of tokens from a larger universe. Minimum qualification (hard rule). Submission must achieve ≥ 75% coverage of the ground-truth subset on the hidden evaluation (i.e., Recall ≥ 0.75). Submissions below this threshold are not eligible for ranking or prizes. Primary metric. Jaccard Index (Intersection-over-Union) between predicted subset and ground truth on the hidden evaluation. Tie-breakers (in order): (1) Higher Precision, (2) Higher F1, (3) Earlier Final Submission timestamp. Public vs. private. If a public leaderboard is shown, it may use a subset of evaluation data; final standings are determined by the private evaluation. 9) Tools & Compute Environment: any runtime is fine. Compute caps: none. Automated tools: allowed only as wrappers that ultimately train CatBoost models (no other learners anywhere in the pipeline). 10) Integrity & Fair Play No collusion or sharing of ground truth, predictions, or private evaluation outcomes across teams. No probing/attack vectors intended to infer hidden labels or reverse-engineer evaluation endpoints. No leakage. Designs must prevent label leakage or hindsight use. Sponsor may request a rerun (e.g., screen-share or containerized execution). Failure to reproduce within reasonable tolerance may result in disqualification. Sponsor may disqualify any Submission that violates the letter or spirit of these Rules. 11) Winner Verification & Deliverables Verification: Winners must complete reasonable verification steps requested by Sponsor (e.g., eligibility confirmations and payment details). Deliverables: Winning teams must provide complete code (training + inference), configs/seeds, and documentation sufficient to fully reproduce their Final Submission. License to Sponsor: As a condition of receiving a prize, winners grant Sponsor a perpetual, worldwide, non-exclusive, royalty-free license to use, reproduce, and evaluate the winning code/models for internal research and productization. 12) Updates & Cancellation Sponsor may post clarifications or updates in Discord. Material changes will be announced reasonably in advance where feasible. Sponsor may cancel or suspend the Competition if fraud, technical failure, or other factors impair integrity or feasibility. 13) Contact Questions and official announcements: https://discord.gg/cabApvps