# Official Competition Rules — AgentEval Part I: Grounded RAG Benchmark ## ENTRY CONSTITUTES ACCEPTANCE **ENTRY IN THIS COMPETITION CONSTITUTES YOUR ACCEPTANCE OF THESE OFFICIAL COMPETITION RULES.** By registering, accessing the data, or submitting an entry, you agree to comply with these Rules, the Competition Website content, and Kaggle’s Foundational Competition Rules. If you do not agree, do not enter or submit. This is a **skills-based, learning-first, community-hosted** competition designed to promote practical evaluation of Retrieval-Augmented Generation (RAG) systems for **grounded answering, citation accuracy, and hallucination control**. **Important:** You may not sign up or participate using multiple Kaggle accounts. One person = one Kaggle account. --- ## 1. COMPETITION-SPECIFIC TERMS ### 1.1 Competition Title **AgentEval Part I: Grounded RAG Benchmark** ### 1.2 Competition Sponsor (Host) **Sweety Seelam** (Community Host) ### 1.3 Competition Sponsor Address **Not applicable (community-hosted).** For platform administration, Kaggle’s platform terms apply. ### 1.4 Competition Website The official Competition page on Kaggle: `https://www.kaggle.com/competitions/agent-eval-part-i-grounded-rag-benchmark` ### 1.5 Prizes **No prizes.** This is a free, unpaid educational competition. (Accordingly, any “Winner License” or “Winner Obligations” in these Rules are **learning-only** and do not involve prize payment.) ### 1.6 Winner License Type **Not applicable (no prizes).** Participants retain ownership of their code/models. Public notebooks and writeups remain subject to Kaggle’s platform terms. ### 1.7 Competition Data Access and Use You may access and use the Competition Data **for the purposes of participating in this competition, learning, education, and non-commercial research** only. You may not use the Competition Data: - to build commercial products, - for private client work, - for paid consulting deliverables, - or to republish the dataset as a standalone dataset elsewhere. --- ## 2. COMPETITION-SPECIFIC RULES ### 2.1 Team Limits - Maximum team size: **3** - Team mergers: allowed if Kaggle supports it for community competitions, and only if you remain within submission limits. ### 2.2 Submission Limits - Maximum **5 submissions per day** - Maximum **2 final submissions** selected for private leaderboard scoring (if enabled by Kaggle settings) ### 2.3 Competition Timeline All dates (start date, final submission deadline, and private leaderboard finalization) are shown on the Competition **Overview → Timeline** section. The Competition Host may adjust dates only through Kaggle’s official settings. --- ## 3. COMPETITION DATA RULES (SECURITY + FAIR USE) ### 3.1 Data Security You agree to use reasonable measures to prevent unauthorized access to Competition Data. You must not: - redistribute the Competition Data as a dataset outside the competition, - post the dataset on another website, - share copies of the dataset with non-participants. ### 3.2 No Reverse Engineering / No Leakage You must not attempt to: - reverse engineer the private evaluation set, - infer hidden answers using leaderboard probing strategies, - share suspected private labels, leakage patterns, or reconstructed gold answers. If you discover potential leakage or evaluation issues, report them in a **private message to the Host** or in a discussion thread without revealing exploit details. --- ## 4. EXTERNAL DATA, MODELS, LLMs, AND TOOLS ### 4.1 External Data External data is allowed only if it is: - publicly available at no cost, and - reasonably accessible to all participants. You must clearly cite external sources used in your notebook/writeup. ### 4.2 External Models / LLM APIs Using open models (local inference) is allowed. Using paid or restricted LLM APIs is allowed only if: - access is broadly available, and - the cost is minimal/reasonable (i.e., does not exclude participants). If you use paid APIs, you must disclose it in your notebook and explain how someone else could reproduce results. ### 4.3 Automated Machine Learning Tools (AutoML) AutoML tools are permitted if you hold an appropriate license and you comply with these Rules. Your submission must still be reproducible and must not violate any data restrictions. --- ## 5. SUBMISSION REQUIREMENTS (FORMAT + VALIDITY) ### 5.1 Required Submission Format Your submission must match the required CSV format described in the Evaluation page, including: - `question_id` - `answer` - `cited_chunk_ids` ### 5.2 Citation Validity - `cited_chunk_ids` must be a list of chunk IDs from `corpus.csv` - Invalid chunk IDs may be ignored or penalized by the metric - Excessive citation (citing too many chunks) may be penalized ### 5.3 Prohibited Content Submissions must not include: - personally identifying information (PII), - harmful or unsafe content, - hateful or discriminatory content, - attempts to exploit platform vulnerabilities. --- ## 6. FAIR PLAY AND PARTICIPANT CONDUCT ### 6.1 One Account Per Person **Multiple Kaggle accounts are not permitted.** If detected, the Host may request Kaggle moderation and disqualify related submissions. ### 6.2 Collaboration and Sharing - Sharing general ideas is encouraged. - Sharing exact private solutions, private labels, or exploit strategies is not allowed. ### 6.3 Reproducibility Expectations (Strongly Encouraged) This is a benchmark-style competition. Participants are encouraged to publish notebooks that: - explain retrieval approach, - show citation strategy, - discuss hallucination mitigation techniques. --- ## 7. DISQUALIFICATION A participant or team may be disqualified for: - violating these Rules, - cheating or leaderboard probing, - using restricted/private data, - attempting to leak or reconstruct private labels, - violating Kaggle’s Foundational Competition Rules. --- ## 8. LIMITATION OF LIABILITY / NO WARRANTY The Competition Data and evaluation are provided “as-is” for learning purposes. The Host makes no warranties regarding accuracy, completeness, or fitness for any purpose. --- ## 9. GOVERNING TERMS This competition is hosted on Kaggle; Kaggle’s platform terms and Foundational Competition Rules apply. Where these Rules conflict with Kaggle’s platform rules, Kaggle’s rules control. --- ## 10. CONTACT For questions, use the competition Discussion board. For suspected leakage or sensitive issues, contact the Host privately.