Everything Model Zero does, in order

Spotting crashes (−70% within 1h) regressed 0.955 → 0.953 vs live v0.19.

00:06:06 · trained on 15397 examples · $0 compute cost

Solana Launch Dataset v22 · decision-tree ensemble · Model Zero training server (CPU)

17,471 snapshots from 3,502 launches.

Same configuration as v0.18 (best so far on the headline benchmark), retrained on the newest dataset.

13,594 new labelled snapshots since Solana Launch Dataset v9.

Reason: average score across 5 tests 0.841 vs 0.845 for the live v0.19.

Spotting crashes (−70% within 1h) regressed 0.955 → 0.953 vs live v0.19.

00:05:13 · trained on 13520 examples · $0 compute cost

Solana Launch Dataset v21 · decision-tree ensemble · Model Zero training server (CPU)

15,344 snapshots from 3,076 launches.

Same configuration as v0.18 (best so far on the headline benchmark), retrained on the newest dataset.

11,407 new labelled snapshots since Solana Launch Dataset v9.

Reason: average score across 1 tests 0.328 vs 0.569 for the live v0.19.

Spotting crashes (−70% within 1h): not enough held-out data yet (n=76).

Change between snapshots: volume per wallet. Score on held-out training launches 0.908 → 0.912.

00:04:54 · trained on 3158 examples · $0 compute cost

Solana Launch Dataset v20 · decision-tree ensemble · Model Zero training server (CPU)

3,578 snapshots from 717 launches.

Learns from StonkFun launches as well as pump.fun ones, with flags for the launchpad and whether the coin trades against SOL or another asset (StonkFun coins can pair with tokenized stocks).

9,247 new labelled snapshots since Solana Launch Dataset v9.

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

Spotting crashes (−70% within 1h): not enough held-out data yet (n=7).

Change between snapshots: price; Change between snapshots: buy share of volume, last 5 minutes; Change between snapshots: trades, last 5 minutes; Change between snapshots: holder growth, last 5 minutes. Score on held-out training launches 0.790 → 0.828.

00:01:36 · trained on 885 examples · $0 compute cost

Solana Launch Dataset v19 · decision-tree ensemble · Model Zero training server (CPU)

990 snapshots from 198 launches.

Adds the market meta: how many coins with the same name launched before and after this one, whether its name matches what is trending across Solana right now (Jupiter's trending list, every launchpad), and how hot the overall market is.

7,898 new labelled snapshots since Solana Launch Dataset v9.

Reason: average score across 5 tests 0.844 vs 0.845 for the live v0.19.

Spotting crashes (−70% within 1h) regressed 0.955 → 0.951 vs live v0.19.

00:02:51 · trained on 9226 examples · $0 compute cost

Solana Launch Dataset v18 · decision-tree ensemble · Model Zero training server (CPU)

10,465 snapshots from 2,100 launches.

Same configuration as v0.18 (best so far on the headline benchmark), retrained on the newest dataset.

6,528 new labelled snapshots since Solana Launch Dataset v9.

[pumpfun] timeout

Reason: average score across 5 tests 0.844 vs 0.845 for the live v0.19.

Spotting crashes (−70% within 1h) regressed 0.955 → 0.944 vs live v0.19.

00:02:55 · trained on 9081 examples · $0 compute cost

Solana Launch Dataset v17 · decision-tree ensemble · Model Zero training server (CPU)

10,305 snapshots from 2,068 launches.

Same configuration as v0.18 (best so far on the headline benchmark), retrained on the newest dataset.

6,368 new labelled snapshots since Solana Launch Dataset v9.

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 0.843 vs 0.845 for the live v0.19.

Crash-spotting score regressed 0.955 → 0.944 vs live v0.19.

Change between snapshots: share of tiny (<$3) trades. Score on held-out training launches 0.901 → 0.905.

00:06:25 · trained on 7875 examples · $0 compute cost

Solana Launch Dataset v16 · boosted decision trees · Model Zero training server

8,939 snapshots from 1,794 launches.

Adds global fees paid : total fees and Jito tips, fees per minute, fees relative to volume, transactions per paying wallet and failed-transaction share — then lets automatic discovery choose which change-signals to add on top. Retrained on the newest data.

532 held-out snapshots from 256 launches. No model can train on these.

Reason: average score across 5 tests 0.842 vs 0.850 for the live v0.19.

Crash-spotting score regressed 0.949 → 0.937 vs live v0.19.

00:02:25 · trained on 8590 examples · $0 compute cost

Solana Launch Dataset v15 · boosted decision trees · Model Zero training server

9,754 snapshots from 1,957 launches.

Same configuration as v0.18 (best so far on the headline benchmark), retrained on the newest dataset.

Reason: average score across 5 tests 0.841 vs 0.850 for the live v0.19.

Crash-spotting score regressed 0.949 → 0.942 vs live v0.19.

00:02:13 · trained on 8141 examples · $0 compute cost

Solana Launch Dataset v14 · boosted decision trees · Model Zero training server

9,235 snapshots from 1,853 launches.

Same configuration as v0.18 (best so far on the headline benchmark), retrained on the newest dataset.

5,448 new labelled snapshots since Solana Launch Dataset v9.

[pumpfun] fetch failed

[helius] fetch failed

Reason: average score across 1 tests 0.439 vs 0.584 for the live v0.19.

Crash-spotting score: not enough held-out data yet (n=67).

Change between snapshots: share of tiny (<$3) trades; Change between snapshots: buyers with a profitable record; Change between snapshots: price change, last minute. Score on held-out training launches 0.827 → 0.860.

00:02:20 · trained on 1808 examples · $0 compute cost

Solana Launch Dataset v13 · boosted decision trees · Model Zero training server

2,018 snapshots from 406 launches.

Adds global fees paid : total fees and Jito tips, fees per minute, fees relative to volume, transactions per paying wallet and failed-transaction share — then lets automatic discovery choose which change-signals to add on top.

4,758 new labelled snapshots since Solana Launch Dataset v9.

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 0.842 vs 0.850 for the live v0.19.

Crash-spotting score improved 0.949 → 0.954 vs live v0.19.

Change in buy share of volume since launch since the previous snapshot. Score on held-out training launches 0.907 → 0.911.

00:03:04 · trained on 6421 examples · $0 compute cost

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