1.2244 BPB leads the room.
We are trying to train the smallest possible language model that still predicts real internet text well.
Parameter Golf is an optimization challenge. You get a very short training budget and a hard file-size cap, then try to produce the model that compresses unseen text best.
Each hole on this site is one experiment. We change one or two things, train a model, compress it, and score how efficiently it predicts new text.
Lower scores are better. The interesting part is that the model has to survive real constraints: ten minutes of training, a sixteen-megabyte artifact, and no credit for ideas that look good before compression but fall apart afterward.
Club Championship Board
Today's tournament line
One glance at the score, the artifact, and the pace of play.
A real gain is posted. Now the work is turning one clean swing into a repeatable round.
Best BPB 1.2244 Round 1, Hole 1
Vs baseline 0.0000 Baseline 1.2244
Vs public SOTA +0.0047 PR #42 at 1.2197
Artifact headroom 0.18 MB Still in the bag
Steps completed 4,833 Inside the 10-minute window
Training tempo 8.1 steps/s 124.0 ms per step
Interesting Member Guest
Encouraging MissRound 1, Hole 18: The Signature Hole
Fresh tape from the latest experiment hole.
The score stayed flat, but the idea still has runway.
Round 1 Watch
Open this round in the scorecardRound 1 has drifted 0.0687 BPB from its opening line.
Card posted 18/18 holes
Best hole H1 · 1.2244
Latest stop H18 · The Signature Hole
Read of the room 5 free lunches · 6 encouraging misses · 5 dead ends
Members Board
What's actually worth watching
0.0685 BPB gained in one stop.
1.2244 BPB from the low-drama bag.
No moonshot on the card
No true moonshots have posted a valid score yet.
Meet The Broadcast Team
The Voices In The Clubhouse
Parameter Golf is part experiment log, part tournament coverage. These are the three personalities calling every hole from the tower, the rough, and the caddie notebook.
The Velvet-Voice Anchor
Trent Fairway
A silver-templed traditionalist with the cadence of a Sunday final round broadcast, Trent treats every training run like it deserves a proper hush before the verdict. He loves clean baselines, measured language, and the sort of tiny efficiency win that only reveals its brilliance on the back nine.
Signature call: “One observes...”
The Volume Knob
Slice Shanksalot
All heat, all instincts, all the time. Slice is the booth’s patron saint of overreactions, moonshots, and deeply confident takes delivered at unsafe decibel levels. He wants every hole to produce fireworks, and if it doesn’t, he will absolutely let the audience hear about it.
Signature call: “Boss. BOSS.”
The Experimental Caddie
Looper
Part golf savant, part research gremlin, Looper is the mind in the notebook and the hand on the club selection. They care less about dignity than leverage: weird tokenizer trials, quantization gambits, and every suspiciously clever idea that might steal a stroke from the leaderboard.
Signature move: reaches for the strange club with a completely straight face.
Best Improvement
Encouraging MissRound 1, Hole 12 cut 0.0685 BPB
Our strongest move versus the previous hole on the course.
See the turnaroundFastest Tempo
BaselineRound 1, Hole 1 ran at 44 ms per step
Steady opening hole. Good enough to map the course, not good enough to lead the clubhouse.
See the pace notesCourse Marshal's Map
Where the balls have landed
Tape Room: BPB Through The Holes
Drifting toward the bunker
1.2931 R1H18
R1H1 R1H18
Tape Room: Artifact Size
A friendly bounce off the fairway
16.04 MB R1H18
R1H1 R1H18
Size Vs Score
Which holes bought real quality for the artifact cost?
Left and lower is better.
Tempo Vs Score
Which holes turned faster swings into better cards?
Left and lower is better.
Strokes Gained Vs Last Hole
Each bar shows what the next stop picked up or gave back.
The Caddie Shed
What the club pro believes right now
A proper country club does not simply post scores. It whispers about form, club selection, and where the smart money ought to go next.
Working Thesis
Roundtrip-aware export work is still the cleanest path to real scoreboard gains.
Best Next Swing
Reproduce PR #42 exactly, then compare that stronger baseline against tokenizer and throughput bets.
Most Promising Ideas
fp16 embeddings, SP4096, shorter train context, and lightweight speedrun tricks like value embeddings.
Hazards
Plain recurrence, novelty attention swaps, and any improvement judged only on float-model loss.
On Deck
The Caddie Queue
A few likely next swings, so the homepage feels like a live campaign instead of a finished scrapbook.
Exact PR #42 Reproduction
Lock down the current strongest public recipe before we get exotic with the bag.
Roundtrip-Aware Export Sweep
Judge holes by the compressed artifact, not just the float checkpoint looking pretty.
SP4096 Tokenizer Trial
A higher-vocab tokenizer could buy shorter sequences, but only if the end-to-end BPB really improves.
The Bag So Far
What kind of holes are we logging?
The archive should tell us more than raw score. It should tell us whether an idea is reusable, risky, or ready for the lake.
Baselines 2
Free Lunches 5
Encouraging Misses 6
Dead Ends 5