Pre-launch· Token not live yet — no rewards or treasury until launchWhat this means
← Model
Automatic signal discovery

TRAIN v0.20

Automatic signal discovery: TRAIN generated candidate signals (how every measure changed since the previous snapshot), tested each on held-out folds of the training launches, and kept only the ones that improved its predictions.
Completed

Definition

Model type
Decision-tree ensemble
Signals used
… + change signals
Technical details
max_depth
2
subsample
0.8
n_estimators
200
learning_rate
0.05
min_samples_leaf
10
Random seed
1020
Config hash
52680eaae2642619
Created
05 Oct 17:21
Completed
05 Oct 17:24
Artifact
models/v0.20/model.json
sha256 bc8d5b927fbf6b81…

Dataset & compute

Dataset
Solana Launch Dataset v12
7,235 snapshots · 1,452 launches · sha256 d45a86d8a29b
Training runs
1
Compute
TRAIN training server (CPU)
GPU hours
—
no GPU used
CPU time
1,165.1 s
Provider cost
$0.00
nothing purchased

Benchmarks

held-out test set · change vs v0.19
Spotting crashes (−70% within 1h)
0.958
Spotting survivors (still traded after 1h)
0.963
Spotting 2× runs (within 1h)
0.854
Calling the 1h direction (down / flat / up)
56.9%
Spotting creator dumps (within 1h)
0.861
Spotting crashes (within 6h)
insufficient data
Spotting survivors (6h)
insufficient data
Calling the 6h direction
insufficient data
Spotting survivors (24h)
insufficient data
Spotting migrations (within 24h)
insufficient data

Validation metrics (during training)

collapse_1h
AUC 0.994
n=814
collapse_6h
—
n=0
reach_2x_1h
AUC 0.926
n=814
survival_1h
AUC 0.962
n=814
survival_6h
—
n=0
survival_24h
—
n=0
migration_24h
—
n=0
trajectory_1h
acc 98.0%
n=814
trajectory_6h
—
n=0
creator_exit_1h
AUC 0.805
n=202

Validation launches are separate from training launches and from the locked test set. Benchmarks above are the public numbers.

Signals this version discovered

45 candidate signals tested on held-out folds of the training launches. Cross-validated AUC 0.907 → 0.911.

What this version relies on

held-out validation launches
Collapse risk · 1h
Volatility, last 5 minutes−0.018
Trades per wallet−0.006
Buy share of volume, last 5 minutes−0.003
Trades, last 5 minutes−0.002
Holder growth, last 5 minutes−0.001
Price change, last 5 minutes−0.000
Survival · 1h
Net SOL into the curve−0.113
Market cap−0.015
Creator's current holding−0.009
Buy share of volume since launch−0.006
Change between snapshots: buy share of volume since launch−0.004
Creator's initial buy−0.004
Reaches 2× · 1h
Holders (from trades)−0.053
Price change, last minute−0.040
Net SOL into the curve−0.026
Holder growth, last 5 minutes−0.013
Top-5 holders' share−0.002
Creator's initial buy−0.002
Creator exit · 1h
Net SOL into the curve−0.073
Creator's current holding−0.071
Age of the launch−0.025
Buy share of volume since launch−0.011
Buy share of volume, last 5 minutes−0.008
Volume acceleration−0.008

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

Training runs

Attempt 1Succeeded
05 Oct 17:21 · 00:03:04 · 6,421 examples · $0.00

Timeline

  1. 05 Oct 17:24
    TRAIN v0.20 not deployed; v0.19 stays live
    Reason: average score across 5 tests 0.842 vs 0.850 for the live v0.19.
  2. 05 Oct 17:24
    Benchmark completed: TRAIN v0.20
    Crash-spotting score improved 0.949 → 0.954 vs live v0.19.
  3. 05 Oct 17:24
    Benchmark started: TRAIN v0.20
  4. 05 Oct 17:24
    TRAIN v0.20 discovered 1 useful new signal
    Change in buy share of volume since launch since the previous snapshot. Score on held-out training launches 0.907 → 0.911.
  5. 05 Oct 17:24
    Training completed: TRAIN v0.20
    00:03:04 · trained on 6421 examples · $0 compute cost
  6. 05 Oct 17:21
    Training started: TRAIN v0.20
    Solana Launch Dataset v12 · decision-tree ensemble · TRAIN training server
  7. 05 Oct 17:21
    TRAIN v0.20 queued
    Automatic signal discovery: TRAIN generated candidate signals (how every measure changed since the previous snapshot), tested each on held-out folds of the training launches, and kept only the ones that improved its predictions.