Before market close, what is the shortest wall-clock time to train an AI that achieves ≥50% zero-shot accuracy on HellaSwag, using a single consumer GPU?

"single consumer GPU" means the most popular GPU in the November 2030 Steam Hardware Survey. If the Steam Hardware Survey becomes defunct or inaccessible, a substitute source in the same spirit will be chosen at my discretion (or the mods, if I am inactive). Training must use only one such GPU with no offloading to other accelerators.

If the training was not performed on the reference GPU, the equivalent wall-clock time is calculated as: (total training FLOPs as reported) ÷ (published bf16 peak FLOPS of the reference GPU without sparsity). Reports that were actually run on the reference GPU resolve on their reported wall-clock time directly.

Will be based on the fastest credible public report (even a detailed blog post with reproducible methodology would count) published before market close. The report must describe the training setup in enough detail to be in principle reproducible. Any dataset and any training tricks are permitted. If no credible report exists, resolves 1000 - 9999+ minutes.