CodeSOTA · Research question · RQ-PG-0001Updated 2026-08-20
What we want to learn

How do we minimize language-model BPB under a 16 MB artifact budget?

Which architecture, training, quantization and test-time adaptation choices minimize FineWeb validation bits per byte while the complete artifact remains under 16,000,000 bytes?

Status
partially answered
Evidence
strong evidence
Graph
5 hypotheses · 5 experiments
Claims
5 public · 5 attachments
Reproductions
0 recorded

The pinned leaderboard moves from the 1.2244 baseline to 1.0565 BPB. The selected #1394–#1530 milestones improve from 1.08563 to 1.07336, but they are a frontier progression rather than one literal code lineage.

Why this question exists.

OpenAI Parameter Golf produced a dense public record of merged experiments, source PRs, declared predecessors, per-seed evidence and reproduction instructions. CodeSOTA imports the official leaderboard as a pinned research dataset.

Why it matters. This is a working model of cumulative empirical research: changes can be traced to their source, score progression can be separated from code ancestry, and every claim can retain the exact upstream evidence.

Keep the supported and refuted branches.

A negative result remains visible research memory. Each branch links its protocol and the claim the evidence supports or refutes.

H-PG-1394supported

The SP8192 recurrence stack improves parameter efficiency

The combined SP8192, GPTQ-embedding, depth-recurrence and SD-clip stack improves FineWeb validation BPB under the Parameter Golf budget.

H-PG-1413supported

QK gain and legal score-first TTT improve the SP8192 stack

Raising QK gain and adding legal score-first test-time training improves the PR #1394 stack under the same artifact and compute constraints.

H-PG-1493supported

Recurrence and parallel residual routing compound

The combined three-layer recurrence, parallel residual routing, QK-gain and legal TTT variant improves the PR #1394 base stack.

H-PG-1514supported

The Muon 0.97 and causal-tilt bundle improves its predecessor

The combined Muon 0.97 and causal token n-gram tilt variant improves the direct PR #1413 predecessor; the score delta is not attributed to either change alone.

H-PG-1530supported

The variable-length attention bundle moves the frontier

The combined variable-length attention, fused MLP and document-independent LoRA TTT variant improves its contemporary Parameter Golf base stack.

What the graph can currently say.

F-PG-0001 · confirmed

Frontier progression is not the same thing as code lineage

The selected milestone scores descend from 1.08563 to 1.07336 BPB, but the declared parents branch: PR #1514 names #1413 as its direct predecessor, while PR #1530 derives from #1523 and incorporates work from #1354 and #1552. Adjacency on a leaderboard must not create a DERIVED_FROM edge.

Visible ownership.

Stable ID
RQ-PG-0001
Created by
CodeSOTA GitHub importer
Contributors
OpenAI Parameter Golf contributors · Kevin Clark · dexhunter · bigbag · samacqua
Visibility
public