LoRA Speedrun leaderboard tracks fine-tuning speed for Qwen2.5-1.5B
July 20, 2026
This leaderboard benchmarks wall-clock fine-tuning speeds on a single L40S GPU using Qwen2.5-1.5B. The current top record achieves 61.1% on GSM8K in 6m 05s by using sequence packing and completion-only loss masking.
HOW THIS AFFECTS YOU
●
builderYou can use these verified optimization techniques like sequence packing to reduce training costs and latency.
●
researcherThe standardized benchmark provides a controlled environment to test hardware-specific fine-tuning efficiency.