LoRA Speedrun
A frozen-task, frozen-hardware public leaderboard for speedrunning LoRA fine-tuning of small language models, with every record independently re-verified on a cloud GPU sandbox.
LoRA Speedrun is a public wall-clock leaderboard for LoRA fine-tuning, modeled on modded-nanogpt's speedrunning format. Each track fixes a task, a target accuracy, and hardware, for example fine-tuning Qwen2.5-1.5B to at least 57% on GSM8K on a single L40S GPU, and ranks entries by training time.
Every record is independently re-run three times with fresh seeds on identical hardware before it counts, using a Modal L40S sandbox; Modal's free monthly compute credits let anyone attempt or re-verify a record with one command. Entries document their technique, such as sequence packing, completion-only loss masking, data pruning, or fused GEMM kernels.
A second track applies the same format to fine-tuning SmolLM2-1.7B on SQuAD. It is written in Python and licensed under MIT.