Post by Kishan Vavdara

Happy to share our team finished 2nd (Public) and 5th (Private) on Kaggle’s Jigsaw - Agile Community Rules Classification Competition —and this pushed me to Competition Master 🏅. We also took home a $5k prize 🎉. The challenge (in one line): Reddit moderation with messy, real-world labels—predict which rule (if any) a comment may have broken. How we tackled it : Fine-tuned LLMs (Llama-3.1 8B, Qwen-3 8B) with DDP + test-time training, using a minimal instruction + rule + comment → constrained binary decoding for comparability, comparable scores. Retrieval + embeddings (FAISS + multi-embedding heads with LR + kNN) to ground predictions in similar historical examples. NLI framing with DeBERTa to reason explicitly about rule violations. vLLM inference (token parallelism, constrained decoding) to keep throughput high. Per-rule rank stacking to blend all signals robustly across rules. Huge thanks to my teammates Benedikt Droste , Ralf Kinkel, Kohki Horie(c-number), Ivan Isaev and to the Kaggle community for excellent discussions and shared baselines. Full write-up: https://lnkd.in/gZn4znXB #kaggle #machinelearning #nlp #llm #moderation