Hackathons: Kaggle (page 4)
Hackathons · Kaggle: 157 entries in total; this page lists entries 46–60.
G-RaR: Rubric-Based Reinforcement Learning for Structured Reasoning in Gemma
Evidence grade: for discovery only; must not by itself support a formal conclusion.
Training Gemma-2-2B for Explicit Reasoning with LoRA SFT and LLM-Judge-Based GRPO on TPU
Evidence use: For discovering how winning projects were implemented, their data resources, evaluation methods and application constraints; specific conclusions must be checked against the original project.
Limitations: Hackathon awards are not equivalent to peer review or independent reproduction. Project descriptions may change with the Kaggle page; check code, data and dependency versions before use.
Link last checked: 2026-09-24
Pinocchio-1B: Creating a Reasoning Model in 3 Acts (SFT → SimPO → GRPO)
Evidence grade: for discovery only; must not by itself support a formal conclusion.
Teaching Gemma 3 1B to "think" using Tunix within a single 9-hour TPU session
Evidence use: For discovering how winning projects were implemented, their data resources, evaluation methods and application constraints; specific conclusions must be checked against the original project.
Limitations: Hackathon awards are not equivalent to peer review or independent reproduction. Project descriptions may change with the Kaggle page; check code, data and dependency versions before use.
Link last checked: 2026-09-24
IDEA-E Distillation with Curriculum Guided GRPO Training
Evidence grade: for discovery only; must not by itself support a formal conclusion.
Distilling structured, ethical reasoning into Gemma-2-2B using the IDEA-E decision making scaffold and a TF-IDF based reasoning reward.
Evidence use: For discovering how winning projects were implemented, their data resources, evaluation methods and application constraints; specific conclusions must be checked against the original project.
Limitations: Hackathon awards are not equivalent to peer review or independent reproduction. Project descriptions may change with the Kaggle page; check code, data and dependency versions before use.
Link last checked: 2026-09-24
Tunix Hackathon
Evidence grade: for discovery only; must not by itself support a formal conclusion.
Tunix fine-tuning
Evidence use: For discovering how winning projects were implemented, their data resources, evaluation methods and application constraints; specific conclusions must be checked against the original project.
Limitations: Hackathon awards are not equivalent to peer review or independent reproduction. Project descriptions may change with the Kaggle page; check code, data and dependency versions before use.
Link last checked: 2026-09-24
Eliciting Reasoning via On-Policy Distillation
Evidence grade: for discovery only; must not by itself support a formal conclusion.
Using an Offline-Trained Teacher and Synthetic Data
Evidence use: For discovering how winning projects were implemented, their data resources, evaluation methods and application constraints; specific conclusions must be checked against the original project.
Limitations: Hackathon awards are not equivalent to peer review or independent reproduction. Project descriptions may change with the Kaggle page; check code, data and dependency versions before use.
Link last checked: 2026-09-24
Gemma2-Deep: Incentivizing Gemma to Reason before Answering
Evidence grade: for discovery only; must not by itself support a formal conclusion.
Reinforcement learning with Group Relative Policy Optimization (GRPO) applied to Gemma2, trained on TPUs using Google Tunix.
Evidence use: For discovering how winning projects were implemented, their data resources, evaluation methods and application constraints; specific conclusions must be checked against the original project.
Limitations: Hackathon awards are not equivalent to peer review or independent reproduction. Project descriptions may change with the Kaggle page; check code, data and dependency versions before use.
Link last checked: 2026-09-24
Visualizing Contested Space to Optimize Defensive Outcomes After the Pass
Evidence grade: for discovery only; must not by itself support a formal conclusion.
We develop methods to visualize contested space, and use those methods to predict YAC, completions, and analyze defensive positioning
Evidence use: For discovering how winning projects were implemented, their data resources, evaluation methods and application constraints; specific conclusions must be checked against the original project.
Limitations: Hackathon awards are not equivalent to peer review or independent reproduction. Project descriptions may change with the Kaggle page; check code, data and dependency versions before use.
Link last checked: 2026-09-24
Path Against The Pass
Evidence grade: for discovery only; must not by itself support a formal conclusion.
This project is motivated by understanding how the interaction between receiver and defender movements influences catch probability.
Evidence use: For discovering how winning projects were implemented, their data resources, evaluation methods and application constraints; specific conclusions must be checked against the original project.
Limitations: Hackathon awards are not equivalent to peer review or independent reproduction. Project descriptions may change with the Kaggle page; check code, data and dependency versions before use.
Link last checked: 2026-09-24
Ghostbusters: Back off man, I'm a Data Scientist!
Evidence grade: for discovery only; must not by itself support a formal conclusion.
Evaluating defender movement while the ball is in the air through hypothetical distributions of ghost defenders - By Lucca Ferraz
Evidence use: For discovering how winning projects were implemented, their data resources, evaluation methods and application constraints; specific conclusions must be checked against the original project.
Limitations: Hackathon awards are not equivalent to peer review or independent reproduction. Project descriptions may change with the Kaggle page; check code, data and dependency versions before use.
Link last checked: 2026-09-24
The Decision Point
Evidence grade: for discovery only; must not by itself support a formal conclusion.
Visualizing Optimal Defensive Positioning
Evidence use: For discovering how winning projects were implemented, their data resources, evaluation methods and application constraints; specific conclusions must be checked against the original project.
Limitations: Hackathon awards are not equivalent to peer review or independent reproduction. Project descriptions may change with the Kaggle page; check code, data and dependency versions before use.
Link last checked: 2026-09-24
Quantifying Pass Play Evolution Through Dynamic Completion Probability
Evidence grade: for discovery only; must not by itself support a formal conclusion.
Quantifying Pass Play Evolution Through Dynamic Completion Probability Measuring What Quarterbacks See And What They Create
Evidence use: For discovering how winning projects were implemented, their data resources, evaluation methods and application constraints; specific conclusions must be checked against the original project.
Limitations: Hackathon awards are not equivalent to peer review or independent reproduction. Project descriptions may change with the Kaggle page; check code, data and dependency versions before use.
Link last checked: 2026-09-24
NeuroState: AI-Powered Embedded Systems IDE.
Evidence grade: for discovery only; must not by itself support a formal conclusion.
Design firmware by speaking — visual FSM editor with Gemini 3 Pro voice assistant, hardware simulation, and multi-language code generation
Evidence use: For discovering how winning projects were implemented, their data resources, evaluation methods and application constraints; specific conclusions must be checked against the original project.
Limitations: Hackathon awards are not equivalent to peer review or independent reproduction. Project descriptions may change with the Kaggle page; check code, data and dependency versions before use.
Link last checked: 2026-09-24
Banana Couture – From Prompt to Factory Floor
Evidence grade: for discovery only; must not by itself support a formal conclusion.
From prompt ? CAD ? tech pack ? suppliers—fast.
Evidence use: For discovering how winning projects were implemented, their data resources, evaluation methods and application constraints; specific conclusions must be checked against the original project.
Limitations: Hackathon awards are not equivalent to peer review or independent reproduction. Project descriptions may change with the Kaggle page; check code, data and dependency versions before use.
Link last checked: 2026-09-24
SIMUL-AI-TOR: Video-to-Reality Physics Engine
Evidence grade: for discovery only; must not by itself support a formal conclusion.
Turn any video into an interactive, scientifically accurate 3D laboratory using Gemini 3 Pro
Evidence use: For discovering how winning projects were implemented, their data resources, evaluation methods and application constraints; specific conclusions must be checked against the original project.
Limitations: Hackathon awards are not equivalent to peer review or independent reproduction. Project descriptions may change with the Kaggle page; check code, data and dependency versions before use.
Link last checked: 2026-09-24
Articulate: The Infinite Canvas That Thinks With You
Evidence grade: for discovery only; must not by itself support a formal conclusion.
A fluid, infinite workspace that liberates creativity - allowing you to think, sketch, and create at the speed of thought
Evidence use: For discovering how winning projects were implemented, their data resources, evaluation methods and application constraints; specific conclusions must be checked against the original project.
Limitations: Hackathon awards are not equivalent to peer review or independent reproduction. Project descriptions may change with the Kaggle page; check code, data and dependency versions before use.
Link last checked: 2026-09-24