OLYMPUS: The Agora — Community companion for OpenAI Parameter Golf. Live leaderboard, compliance engine, technique map, funding transparency, compute guide.
-
Updated
May 31, 2026 - Python
OLYMPUS: The Agora — Community companion for OpenAI Parameter Golf. Live leaderboard, compliance engine, technique map, funding transparency, compute guide.
HYDRA · parameter-golf fork — tiny LM ≤16MB / ~10min · ~1.061 BPB verified local · receipts + writeup
Beating OpenAI's 8×H100 Parameter Golf baseline (1.2019 vs 1.2244 bpb) on a laptop, under a 16MB limit
OpenAI Parameter Golf — SOTA 0.3958 BPB (3-seed mean). Beat prior best (0.4416) by 0.0458. Causal BackoffNgramMixer. PR #1094.
Parameter Golf v3: ~1.151 BPB in 16MB — Int6 QAT, BigramHash, SmearGate, SWA, zstd-22
CPU-only language model optimization under a 16 MB constraint — calibration, Bayesian n-gram experts, test-time training, and compression.
Training LLMs under extreme 16MB memory constraints using custom optimizers in PyTorch. tested on H100 clusters (RunPod) for the 16MB OpenAI Parameter Golf Challenge.
Exploratory compact LM research branch derived from LUMI-Arch, focused on structure-sensitive generalization under tight model constraints.
A practical starter repo and working setup for the OpenAI Parameter Golf competition
OpenAI Parameter Golf experiments for parameter-constrained language modeling, ablations, and efficient architecture tradeoffs.
Sparse I/O bottleneck language model using C. elegans connectome wiring + JEPA + SigREG. Parameter Golf submission: 1.79 BPB in a 2MB artifact.
Study fork of openai/parameter-golf: the upstream checkout plus PR-only submissions, provenance records and a curated reading guide to the 16MB language-model leaderboard.
To associate your repository with the parameter-golf topic, visit your repo's landing page and select "manage topics."