TunexRL - Reinforcement Learning for Reliable LLM Explanations using Tunix скачать в хорошем качестве

TunexRL - Reinforcement Learning for Reliable LLM Explanations using Tunix

This video presents my submission for the Google Tunix Hackathon, where I fine-tune Gemma-3-1B to produce explicit, judge-visible reasoning traces using Tunix, Google’s JAX-native post-training library. The goal of this project is not to maximize benchmark accuracy, but to train a model that reliably explains its reasoning in a strict, reproducible format The model is trained end-to-end using GRPO (Group Relative Policy Optimization) in a single Kaggle TPU session, with no inference-time post-processing or output repair. If the model produces invalid output, it is surfaced exactly as generated—matching judge evaluation behavior.