Deeper decision trees — retrained on new data

Model Zero v0.09

Same configuration as v0.05 (best so far on the headline benchmark), retrained on the newest dataset.

Completedpreviously deployed

Definition

Model type

Boosted decision trees

Signals used

Market signals

Technical details

max_depth

subsample

0.8

n_estimators

300

learning_rate

0.03

min_samples_leaf

Random seed

1009

Config hash

da2b7258b9e789f8

Created

05 Oct 09:12

Completed

05 Oct 09:13

Artifact

models/v0.09/model.json

sha256 13382a07e62e622d…

Dataset & compute

Dataset

Solana Launch Dataset v8

3,757 snapshots · 753 launches · sha256 7a3cc82a36d5

Training runs

Compute

Model Zero training server (CPU)

GPU hours

—

no GPU used

CPU time

17.5 s

Provider cost

$0.00

nothing purchased

Benchmarks

held-out test set · change vs v0.08

Spotting crashes (−70% within 1h)

10,012 tests · likely 94–95 · guessing: 50 · ↑ +0.6 pts

94 / 100

Spotting survivors (still traded after 1h)

10,012 tests · likely 87–90 · guessing: 50 · ↓ −0.5 pts

89 / 100

Spotting 2× runs (within 1h)

10,012 tests · likely 82–85 · guessing: 50 · ↑ +0.3 pts

84 / 100

Calling the 1h direction (down / flat / up)

10,012 tests · likely 52–54 · guessing: 33 · ↓ −0.4 pts

53%

Spotting creator dumps (within 1h)

2,471 tests · likely 79–82 · guessing: 50 · ↓ −1.5 pts

80 / 100

Spotting crashes (within 6h)

9,940 tests · likely 50–50 · guessing: 50 · ±0

50 / 100

Spotting survivors (6h)

Validation metrics (during training)

collapse_1h

AUC 0.993

n=425

collapse_6h

—

n=0

reach_2x_1h

AUC 0.979

n=425

survival_1h

AUC 0.897

n=425

survival_6h

—

n=0

survival_24h

—

n=0

migration_24h

—

What this version relies on

held-out validation launches

This version was trained before Model Zero started measuring which signals each model relies on.

For each signal, how much the score drops when that signal is scrambled across launches. Model Zero learns these weights itself from the data; nobody hand-codes them.

Training runs

Attempt 1Succeeded

05 Oct 09:12 · 00:00:13 · 3,332 examples · $0.00

Timeline

11 Oct 12:18
Model Zero v0.09 not deployed; v0.40 stays live
Reason: average score across 9 tests 0.648 vs 0.843 for the live v0.40.
05 Oct 09:13
Model Zero v0.09 deployed — now live
average score across 5 tests 0.812 vs 0.795 for the live v0.08
05 Oct 09:13
Benchmark completed: Model Zero v0.09
Crash-spotting score: 0.911 (n=158).
05 Oct 09:13
Benchmark started: Model Zero v0.09
05 Oct 09:13
Training completed: Model Zero v0.09
00:00:13 · trained on 3332 examples · $0 compute cost
05 Oct 09:12
Training started: Model Zero v0.09
Solana Launch Dataset v8 · boosted decision trees · Model Zero training server
05 Oct 09:12
Model Zero v0.09 queued
Same configuration as v0.05 (best so far on the headline benchmark), retrained on the newest dataset.