LoRA Speedrun – a public wall-clock leaderboard for fine-tuning techniques
·2 min read
Compiled by KHAO Editorial
— aggregated from 1 source + 4 references discovered via search.
See llms.txt for citation guidance.
◌ Single Source
This is modded-nanogpt for fine-tuning: a frozen task, frozen hardware, and a public leaderboard of wall-clock records.
Key facts
Any 24 GB+ card runs the baseline for fast experimenting, bash scripts/setup_gpu.sh, then python harness/run_submission.py submissions/000-baseline --runs 1
How fast can you LoRA-fine-tune Qwen2.5-1.5B to ≥ 57% on GSM8K, on a single L40S
Current record: 6m 05s by, Sequence packing + completion-only loss masking, 2 epochs
Taken so far: sequence packing + completion-only masking (record #1)
Summary
How fast can you LoRA-fine-tune Qwen2.5-1.5B to ≥ 57% on GSM8K, on a single L40S? Attempting and verifying are free: official timing runs on a Modal L40S sandbox, and
Modal's free monthly compute credits cover full runs, so anyone can compete, and anyone
can re-verify any record with one command. Current record: 6m 05s by, Sequence packing + completion-only loss masking, 2 epochs. Full history with verification reports: records/RECORDS.md.