Benedict.

Back to posts

The Next Parameter Golf Run

7 min read
contents

Sections

If I took a clean run at this, I would spend less time hunting for the clever architecture and more time making the bottleneck visible earlier.

That is the cleanest postmortem.

The challenge started as a way for me to learn transformer architecture. It did that. But the stronger lesson was broader: small-model work is only partly about the block diagram. The rest is measurement, compression, evaluation, and research hygiene.

The next run would start there.

If I had to answer the opening question now, I would put it this way:

When the model has nowhere to hide, the move I trust is not just a clever block. It is knowing which part of the pipeline actually moved the score.

That is the synthesis the next run should test.

What I Would Keep#

The pieces I would keep are mostly boring, which is a good sign.

PracticeWhy I would keep itFailure mode it prevents
Score the round-tripped artefactThe submitted file is the productCelebrating a checkpoint that compression breaks
Track pre-quant and post-quant BPBCompression damage is part of the model-design problemTreating tensor fragility as bookkeeping
Separate training from evaluationSliding-window scores and better-trained models differCalling an evaluation win a model win
Attach every number to a sourceLogs, metadata, READMEs, and git history have different weightTurning stale prose into public evidence
Use agents for evidence passesThey help keep broad comparisons organisedLetting one confident summary flatten ambiguity

That is the base layer for another run.

What I Would Be More Careful About#

I would be much more careful about claims that mix mechanisms.

For example, a run might change context length, learning rate, batch size, optimiser momentum, and evaluation method. If the score improves, the result is useful, but it does not tell me which lever mattered unless the ablations are clean.

That is fine for a competition. It is weaker for learning.

The next run would separate two modes:

ModeGoalSuccess condition
Record chasingImprove the final scoreStack whatever works under the rules
UnderstandingIsolate a mechanismChange one thing and measure the right gap

Both modes are valid. The mistake is pretending they answer the same question.

If I am writing publicly, I need to mark which mode produced a claim. “This won” is not the same as “this explains why it won”.

Experiments I Would Run First#

My next pass would start with a small set of ranked hypotheses.

RankHypothesisControlsPrimary metricDecision rule
1Some tensors deserve more bytesFixed checkpoint; vary only tensor precision/compression policyPost-quant BPB delta per byte spentProtect tensors that buy the most BPB per byte; demote the rest
2Evaluation context is a separate leverFixed trained model; vary stride and eval sequence length within the eval budgetBPB plus evaluation wallclockKeep the richest legal context that improves BPB, but label it as evaluation gain
3Longer training context can beat more updatesSame tokenizer, model shape, and evaluation protocol where possible; compare 1024, 2048, 4096Post-quant BPB, steps completed, wallclockTreat it as training signal only if it survives fixed evaluation
4Extra depth and MLP bytes have different payoffCompare depth and MLP size under the same compression policyFinal BPB, artefact bytes, post-quant gapSpend bytes where final BPB improves after roundtrip
5Test-time adaptation has marginal value after eval controlsLock document isolation and stride first; compare no-LoRA vs LoRABPB improvement and eval-time costKeep adaptation only if it beats the fixed evaluation baseline within budget
6Better records improve learning speedEvery run gets source of metric, code diff, hardware, artefact size, and claim statusCompleteness of the claim ledgerNo public claim until the evidence row is filled

This list is less exciting than “try a wild new architecture”. That is the point.

I would rather make the bottleneck visible before trying to beat it.

That is the builder lesson I would carry outside the challenge too: instrument the pipeline before chasing cleverness. Otherwise the clever idea might work, and I still would not know why.

The Result I Would Reproduce Before Reusing#

The Warmdown/Int6 MLP3x result is the obvious example.

It is interesting because the log and submission metadata point to a very strong score, but the folder documentation does not cleanly match. Git history made the public status look unsettled, so I would keep it as an experiment seed, not as a settled claim.

The next run would isolate it:

  • reproduce the exact final score
  • confirm the artefact byte count
  • separate normal roundtrip BPB from sliding-window BPB
  • compare against the same model without the extra MLP
  • document why it was removed or whether it was simply superseded

Until then, it belongs in the category of “interesting, not fully trusted”.

That category is useful. It stops me from either discarding the idea or overclaiming it.

What I Can Now Say#

The scoped claim is narrower than mastery and more useful: I can now separate model quality, artefact quality, evaluation effects, and evidence quality more clearly than when I started.

Before this series, I might have looked at a leaderboard jump and asked:

What architecture trick made the model better?

Now I would ask a different set of questions:

QuestionWhy it matters
Did the raw model improve?Separates learning from packaging
Did the compressed artefact hold up?Separates checkpoint quality from submitted quality
Did evaluation context change?Separates model quality from scoring protocol
Is the number log-backed or prose-backed?Separates evidence from summary
Was the run a clean ablation or a record bundle?Separates explanation from competition progress

That is the skill I wanted from Parameter Golf. Not a certificate. A better instinct for where to look when a small model improves, fails, or lies through its metric.

What I Took From The Process#

The bigger personal takeaway is about learning with agents.

I do not think there is anything noble about refusing to use coding agents or AI tools just to prove that every line was handwritten. That feels like fighting the wrong battle. The more useful standard is whether I still understand the thing I am building, testing, or claiming.

That frame fits how this project changed me. The agents helped with execution, synthesis, comparison, and debate. They made it easier to pick up a difficult area quickly. But the understanding still had to be mine. I had to decide which sources counted, which claims were overreaching, and which tradeoffs were actually worth learning from.

That is the version of agentic work I want to keep using: not delegating the understanding, but using the tools to get to the understanding faster.

If I run Parameter Golf again, the goal is not just a lower BPB. It is a lower BPB with a cleaner explanation.

That is the part that feels most portable outside this challenge. Better systems do not only produce better numbers. They make it easier to tell why the number moved, what tradeoff paid for it, and whether I should trust it enough to build the next thing on top.