Ty Pham-Swann
Pages
CactusBench
Measuring frontier models' ability to carry out quality-assurance tasks on real scientific data using complex visual reasoning.
Post-training a 27B model to frontier performance
Post-training an open Qwen3.6-27B model to beat frontier models on CactusBench with SFT and RLVR.
LostBench
Measuring frontier models on long-horizon spatial reasoning tasks.
Saguaro Phenology
A seven-year, model-labeled dataset of saguaro bloom cycles, built with biologists at Saguaro National Park.
Rescultaya
A metal-sculpture business I built and scaled in high school (2018–2021).