SYNTH LABORATORIES · OPEN RESEARCH

Submissions
leaderboard.

Small-model classification hill climbing with fixed data, sealed evaluation, exact labels, and reproducible receipts.

BANKING77 · DEVELOPMENT

400 fixed examples

● ACTIVE
#SubmissionAccuracyMacro-F1
01Tinker SFT + CISPO v5GPT-OSS-20B · 64 rollouts · 50 updates87.75%86.78%02Tinker SFT v1GPT-OSS-20B · rank-16 LoRA87.50%86.29%

Development rankings guide iteration. Sealed-heldout rankings are kept separate.

BANKING77 · SEALED HELDOUT

Promotion results

400 EXAMPLES
#SubmissionAccuracyMacro-F1
01Tinker SFT v1GPT-OSS-20B51.25%53.95%02GEPA large v5 frontierGPT-OSS-20B44.75%48.19%03BaseGPT-OSS-20B35.25%40.10%

NEW SUBMISSION

Start a climb in Workshop.

Use the ready-made prompt to choose a parent, implement one bounded intervention, and prepare a reproducible development submission.

Open Workshop prompt →Submission template