How fast can you LoRA‑fine‑tune Qwen2.5-1.5B to ≥57% on GSM8K with a single L40S? Modded‑nanogpt provides a frozen task, frozen hardware, and a public wall‑clock leaderboard. Every record is independently re‑run 3× with fresh seeds on identical hardware before it counts, ensuring fairness and reproducibility. The current record—6m 05s by @Saivineeth147—was set using sequence packing and completion‑only loss masking over 2 epochs, achieving 61.1% accuracy.

The task is strict: use only the GSM8K train split, train a PEFT adapter ≤30 M params, and report wall‑clock time on a Modal‑sandboxed L40S. The leaderboard is fully auditable; submissions are verified by CI, automated security checks, and three fresh runs in a network‑blocked environment. Anyone can compete for free using Modal’s monthly credits, and the code, records, and verification reports are all public.

This initiative mirrors the nano GPT speedrun, providing a level playing field for exploring LoRA techniques—rank‑adaptive methods, QLoRA, NeFTune, and more—without the noise of differing models or hardware. It’s a science experiment that rewards the fastest, most efficient fine‑tuning.

Join the race, submit a faster run, and earn a spot on the leaderboard.