filipe(@filicroval):NVIDIA post-trained a model from 54.41% to 93.35% with two prompts Using Codex and TAO agent skills, Cosmos 3 Nano went from 54.41% to 87.14% accuracy on 4-choice Woven Traffic Safety video QA. The agent inspected the annotations, ran the baseline, generated the LoRA config, launched training and evaluated the adapter. A second prompt launched an AutoML sweep and pushed validation accuracy to 93.35%. The first LoRA run used 8 A100s for about 30 minutes of training. The final result required 43 parallel trials across multiple A100 nodes for 19.5 hours.

filipe

@filicroval

data eng | 1xAsus Ascent GX10 | benchmarking local models so you don't have to

Joined May 2023

214 Following 153.4K Followers

filipe@filicroval

2026.07.19 12:38

NVIDIA post-trained a model from 54.41% to 93.35% with two prompts Using Codex and TAO agent skills, Cosmos 3 Nano went from 54.41% to 87.14% accuracy on 4-choice Woven Traffic Safety video QA. The agent inspected the annotations, ran the baseline, generated the LoRA config, launched training and evaluated the adapter. A second prompt launched an AutoML sweep and pushed validation accuracy to 93.35%. The first LoRA run used 8 A100s for about 30 minutes of training. The final result required 43 parallel trials across multiple A100 nodes for 19.5 hours.