Everything TRAIN does, in order

Reason: average score across 5 tests 0.797 vs 0.812 for the live v0.09.

Crash-spotting score improved 0.911 → 0.929 vs live v0.09.

00:00:04 · trained on 3502 examples · $0 compute cost

Solana Launch Dataset v9 · decision-tree ensemble · TRAIN training server

Same configuration as v0.04 (best so far on the headline benchmark), retrained on the newest dataset.

Recorded as an outcome label (never used as a feature).

Recorded as an outcome label (never used as a feature).

Recorded as an outcome label (never used as a feature).

Reason: average score across 5 tests 0.800 vs 0.812 for the live v0.09.

Crash-spotting score improved 0.911 → 0.918 vs live v0.09.

00:00:04 · trained on 3502 examples · $0 compute cost

Solana Launch Dataset v9 · decision-tree ensemble · TRAIN training server

Same configuration as v0.04 (best so far on the headline benchmark), retrained on the newest dataset.

Recorded as an outcome label (never used as a feature).

Reason: average score across 5 tests 0.787 vs 0.812 for the live v0.09.

Crash-spotting score regressed 0.911 → 0.884 vs live v0.09.

00:00:04 · trained on 3502 examples · $0 compute cost

Solana Launch Dataset v9 · decision-tree ensemble · TRAIN training server

Same configuration as v0.04 (best so far on the headline benchmark), retrained on the newest dataset.

Recorded as an outcome label (never used as a feature).

Recorded as an outcome label (never used as a feature).

Reason: average score across 5 tests 0.794 vs 0.812 for the live v0.09.

Crash-spotting score regressed 0.911 → 0.909 vs live v0.09.

00:00:04 · trained on 3502 examples · $0 compute cost

Solana Launch Dataset v9 · decision-tree ensemble · TRAIN training server

Same configuration as v0.04 (best so far on the headline benchmark), retrained on the newest dataset.

Reason: average score across 5 tests 0.795 vs 0.812 for the live v0.09.

Crash-spotting score regressed 0.911 → 0.907 vs live v0.09.

00:00:04 · trained on 3502 examples · $0 compute cost

Solana Launch Dataset v9 · decision-tree ensemble · TRAIN training server

3,937 snapshots from 789 launches.

Same configuration as v0.04 (best so far on the headline benchmark), retrained on the newest dataset.

average score across 5 tests 0.812 vs 0.795 for the live v0.08

Crash-spotting score: 0.911 (n=158).

00:00:13 · trained on 3332 examples · $0 compute cost

Solana Launch Dataset v8 · decision-tree ensemble · TRAIN training server

158 held-out snapshots from 76 launches. No model can train on these.

3,757 snapshots from 753 launches.

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

Recorded as an outcome label (never used as a feature).

average score across 2 tests 0.607 vs 0.602 for the live v0.04

1h trajectory · balanced accuracy regressed 0.485 → 0.420 vs live v0.04.

00:00:11 · trained on 2752 examples · $0 compute cost

Solana Launch Dataset v7 · decision-tree ensemble · TRAIN training server

3,122 snapshots from 626 launches.

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

Recorded as an outcome label (never used as a feature).

Recorded as an outcome label (never used as a feature).

Recorded as an outcome label (never used as a feature).

Reason: average score across 2 tests 0.558 vs 0.602 for the live v0.04.

1h trajectory · balanced accuracy: 0.390 (n=94).

00:00:03 · trained on 2187 examples · $0 compute cost

Solana Launch Dataset v6 · decision-tree ensemble · TRAIN training server

94 held-out snapshots from 45 launches. No model can train on these.

2,492 snapshots from 500 launches.

Same configuration as v0.04 (best so far on the headline benchmark), retrained on the newest dataset.

Recorded as an outcome label (never used as a feature).

average score across 1 tests 0.500 vs 0.333 for the live v0.03

average score across 1 tests 0.333 vs 0.333 for the live v0.01

Reason: .

A bug kept earlier versions from being scored on newly locked test sets, so every deploy check since v0.02 found no shared benchmarks. After the fix, all versions were scored on Held-out test set v4 and the deploy rule was re-run in version order.

58 held-out snapshots from 28 launches. No model can train on these.

58 held-out snapshots from 28 launches. No model can train on these.

Reason: no shared benchmarks yet; keeping the live model.

Crash-spotting score: not enough held-out data yet (n=220).

00:00:03 · trained on 679 examples · $0 compute cost

Solana Launch Dataset v5 · decision-tree ensemble · TRAIN training server

220 held-out snapshots from 44 launches. No model can train on these.

784 snapshots from 157 launches.

Adds the realized track record (inside TRAIN's sample, before the snapshot) of the wallets buying the launch — how many are known, profitable, losing or high-frequency.

Recorded as an outcome label (never used as a feature).

Recorded as an outcome label (never used as a feature).

Recorded as an outcome label (never used as a feature).

Reason: no shared benchmarks yet; keeping the live model.

Crash-spotting score: not enough held-out data yet (n=65).

00:00:04 · trained on 1025 examples · $0 compute cost

Solana Launch Dataset v4 · decision-tree ensemble · TRAIN training server

65 held-out snapshots from 13 launches. No model can train on these.

1,170 snapshots from 235 launches.

Deeper trees (depth 3), more of them (300) with a lower learning rate.

Reason: no shared benchmarks yet; keeping the live model.

No locked test set yet (not enough held-out launches with observed outcomes). Benchmarks will be computed when it is locked.

00:00:01 · trained on 594 examples · $0 compute cost

Solana Launch Dataset v3 · decision-tree ensemble · TRAIN training server

694 snapshots from 139 launches.

Switches to gradient-boosted trees (depth 2, 120 trees), which can learn thresholds and interactions a linear model cannot.

Reason: no shared benchmarks yet; keeping the live model.

No locked test set yet (not enough held-out launches with observed outcomes). Benchmarks will be computed when it is locked.