Phương Nguyễn

NVIDIA Nemotron Reasoning Challenge

Kaggle competition on few-shot rule induction across 6 reasoning problem types (bit manipulation, physics gravity, unit conversion, numeral systems, symbol transform, encryption ciphers). Approach: a solver-distilled curriculum of verified teacher examples used to fine-tune NVIDIA's Nemotron-3-Nano with LoRA. The physics-gravity generator was not observed in the public competitor notebooks reviewed.

Built deterministic per-type solvers and generated 21,311 curated training records across 5 of the 6 problem types. Round-trip re-verification caught errors in the synthetic generators. A rank-32 LoRA training pipeline was implemented for the 30B model, targeting Kaggle's vLLM inference backend.