CodeSOTA · Experiment · E-PG-1413Portable protocol · public evidence graph
Why this run exists

PR #1413 · SP8192 + QK-Gain 5 + Legal Score-First TTT

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

Experiment
E-PG-1413
Status
completed
Outcome
supported
Forks
0 public or local starts recorded
Owner
dexhunter

Method before metrics.

On PR #1413: QK-Gain 5.0 + legal score-first TTT on the PR #1394 SP8192 stack Declared changes include: No SLOT, no pre-quant TTT, no ETLB, no n-gram cache, no tokenizer change. Score-first TTT follows the PR #549 precedent — every chunk is scored under inferencemode() before any parameter update. Same base stack as PR #1394: sp8192 BPE, 11L/512d/8H/4KV, MLP 4×, Partial RoPE 16d, depth recurrence (loop layers 4–5 twice from 50% training), MuonEq-R WD=0.085, full-Hessian GPTQ int6 + int8 embeddings + SD-clip, Brotli+byte-shuffle, EMA. Two changes: (1) QKGAININIT raised from 4.0 → 5.0; (2) added a legal score-first TTT sliding pass (LR=0.005, 3 epochs, freezeblocks=0) as an additional eval mode. Condition 3 (Score before update): Every TTT chunk is scored under torch.inferencemode() BEFORE any parameter update. Training on a chunk only happens AFTER its scoring has been accumulated into losssum. Matches the PR #549 pattern.

Hypothesis
H-PG-1413 · Raising QK gain and adding legal score-first test-time training improves the PR #1394 stack under the same artifact and compute constraints.
Directly extended by
PR #1514 · SP8192 + Muon 0.97 + Legal Score-First TTT
experiment:E-PG-1514

The comparison contract.

A reproduction matches this protocol. A fork changes it and declares the deviation.

model Family
Parameter Golf language model
dataset
FineWeb validation
benchmarks
FineWeb validation BPB
controlled Variables
16,000,000-byte artifact cap · 600-second training budget · 8×H100 SXM evaluation hardware · tokenizer-agnostic BPB
independent Variable
combined implementation changes declared in the source PR
seeds
0 · 42 · 1234
hardware
8×H100 SXM
pr number
1,413
source pr
https://github.com/openai/parameter-golf/pull/1413
source commit
cd04a8b7b57380ea48de4e195ebb65ee27076313
source dataset commit
f5c079314c4877fbb0af378c0abade5a8ca33d3a
seed count
3
metric direction
lower_is_better
artifact limit bytes
16,000,000
training budget seconds
600
declared changes
No SLOT, no pre-quant TTT, no ETLB, no n-gram cache, no tokenizer change. Score-first TTT follows the PR #549 precedent — every chunk is scored under inferencemode() before any parameter update. · Same base stack as PR #1394: sp8192 BPE, 11L/512d/8H/4KV, MLP 4×, Partial RoPE 16d, depth recurrence (loop layers 4–5 twice from 50% training), MuonEq-R WD=0.085, full-Hessian GPTQ int6 + int8 embeddings + SD-clip, Brotli+byte-shuffle, EMA. · Two changes: (1) QKGAININIT raised from 4.0 → 5.0; (2) added a legal score-first TTT sliding pass (LR=0.005, 3 epochs, freezeblocks=0) as an additional eval mode. · Condition 3 (Score before update): Every TTT chunk is scored under torch.inferencemode() BEFORE any parameter update. Training on a chunk only happens AFTER its scoring has been accumulated into losssum. Matches the PR #549 pattern. · No pre-quant TTT on val data. · No tokenizer change — uses PR #1394's SentencePiece BPE 8192 unchanged.
declared relations
[object Object]

What this experiment produced or evaluated.

Produced
SP8192 + QK-Gain 5 + Legal Score-First TTT
external-run:ER-PG-1413
Uses
benchmark:fineweb-validation-bpb

Change the evidence, not just the discussion.

Start locally, inspect the portable manifest, and sign in only if you choose to publish.

Attach a run to this protocol.

Record the measurement and its provenance. New evidence is unverified and pending review until CodeSOTA checks it.

Stable, portable, attributable.

Stable ID
E-PG-1413
Visibility
public
Created
2026-04-06
Updated
2026-08-20