v1.1 — Track 2 live: SQuAD · SmolLM2-1.7B (techniques must transfer)

Track 2 is live, frozen, and certified — a direct response to launch feedback that a single model + task risks overfitting techniques to one setup.

The two tracks

Track	Task · Model	Target	Record to beat
1	GSM8K · Qwen/Qwen2.5-1.5B	≥ 57.0%	6m 05s (packing + completion-only masking)
2	SQuAD v1.1 · HuggingFaceTB/SmolLM2-1.7B	≥ 75.5% EM	11m 08s (naive baseline — wide open)

Deliberately different model families and task types: a technique only proves general by winning on both. Same hardware (Modal L40S sandbox), caps (≤30M adapter params), and 3-fresh-seed verification everywhere. Submissions declare track: t1 or track: t2 in config.

Track 2 details

• Target frozen at tag spec-t2-frozen, calibrated by the same protocol as Track 1 (baseline observed 77.5% EM − ~2pts).
• Baseline verified 3/3 fresh seeds: 77.60 / 77.62 / 77.38% EM. One run's wall-clock was inflated by platform storage I/O — disclosed in the verification report and tracked as a protocol observation in #7.
• The tricks that took −49% on Track 1 (sequence packing, completion-only masking) are unclaimed on Track 2 — likely an easy first record for whoever ports them.

Attempting and re-verifying remain free (Modal monthly credits). Rules: TASK.md · CONTRIBUTING.md

v1.0 — Track v1 frozen, two verified records

First public release of the LoRA Speedrun.

Track v1 (gsm8k-qwen2.5-1.5b-l40s) is frozen — base model, dataset, eval protocol, hardware, constraints, and target are locked (tag `spec-v1-frozen`). Times are comparable from here on.

Verified records so far

#	Technique	Train time	GSM8K
0	Baseline — plain LoRA, no tricks	11m 57s	59.4%
1	+ sequence packing + completion-only masking	6m 05s	61.1%

Both were re-run 3× with fresh seeds on identical hardware (Modal L40S, network-blocked sandbox). Full verification reports are in records/verifications/.

The task

Fine-tune Qwen/Qwen2.5-1.5B to ≥57% on GSM8K using only the GSM8K train split, on one L40S. Score = training wall-clock. Lowest wins. ≤30M adapter params, base frozen.

Attempting and re-verifying any record is free (Modal's monthly credits cover full runs).

Beat the record

Current record to beat: 6m 05s. Open lanes with no record yet: 1-epoch schedules, data pruning, QLoRA, torch.compile, custom kernels. See the pinned "open lane" issues and CONTRIBUTING.md.