Model

TRAIN v0.19

TRAIN is not one big chatbot. It is a set of small, specialised models that each answer one question about a launch from numbers that were observable at the time, plus a layer that turns those answers into plain language.

Current version

Version

v0.19

Status

Live

Architecture

Gradient-boosted decision trees (one ensemble per question)

What it is

Launch structure, deeper trees

Signals used

Market + wallets + launch structure

Deployed

05 Oct 17:21

Training dataset

Solana Launch Dataset v11

7,225 snapshots · 1,450 launches

Same signals with deeper trees, to learn interactions such as 'bundle still holding while snipers exit'.

Architecture

1Snapshot. A launch is described at a fixed age (1, 5, 15, 30 or 60 minutes) by 36 numbers: price path, volume, buy/sell flow, wallets, estimated holder concentration, creator behaviour and history, socials — plus the track record of the wallets buying it (6 signals), launch structure (bundles, snipers, dev history), global fees paid, and how each of these changed since the previous snapshot.
2Structured market model. One head per question (10 heads) turns the snapshot into a probability. Architecture changes between versions; the questions stay fixed so versions are comparable.
3Historical comparables. The nearest launches at the same age in the dataset, and what happened to them.
4Explanation layer. Turns the numbers into sentences. It can only use numbers produced by steps 1–3.

Benchmarks

532 held-out snapshots from 256 launches TRAIN never trained on

Spotting crashes (−70% within 1h)

0.955

—

n=532 · 0.91–0.99

Spotting survivors (still traded after 1h)

0.957

—

n=532 · 0.91–0.99

Spotting 2× runs (within 1h)

0.866

—

n=532 · 0.78–0.93

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

56.9%

—

n=532 · 50–64

Spotting creator dumps (within 1h)

0.880

—

n=112 · 0.81–0.94

Spotting crashes (within 6h)

insufficient data

—

n=0

Spotting survivors (6h)

insufficient data

—

n=0

Calling the 6h direction

insufficient data

—

n=0

Spotting survivors (24h)

insufficient data

—

n=0

Spotting migrations (within 24h)

insufficient data

—

n=0

AUC: the chance the model ranks a random launch that had the outcome above one that didn't. 0.5 is a coin flip, 1.0 is perfect. Balanced accuracy averages the hit rate over each class (chance ≈ 33% for three classes).

Ranges are 90% bootstrap intervals over launches. Benchmarks with too few examples on either side show “insufficient data” instead of a number.

Not benchmarked yet

Holder quality

TRAIN already tracks the record of the wallets buying a launch; a holder-quality outcome that can be measured fairly still needs to be defined.

Distribution health

Needs historical holder snapshots; providers only expose current holders, which would leak the future into past examples.

What the live model relies on

v0.19 · measured on held-out validation launches

Collapse risk · 1h

Volatility, last 5 minutes−0.011

Trades per wallet−0.008

Trades, last 5 minutes−0.005

Net SOL into the curve−0.003

Creator's current holding−0.003

Survival · 1h

Net SOL into the curve−0.144

Creator's current holding−0.013

Buy share of volume since launch−0.012

Volume acceleration−0.006

Volume since launch−0.005

Reaches 2× · 1h

Holders (from trades)−0.046

Price change, last minute−0.029

Net SOL into the curve−0.028

Creator's current holding−0.011

Holder growth, last 5 minutes−0.009

Creator exit · 1h

Creator's current holding−0.105

Net SOL into the curve−0.049

Age of the launch−0.026

Buy share of volume since launch−0.018

Website set at launch−0.016

Model history

Spotting crashes (−70% within 1h)

v0.010.450 – 0.961v0.25

Versions are never overwritten or deleted, including failed ones and ones that got worse.

Training datasets

Solana Launch Dataset v17

10,305 snapshots

2,068 launches · 05 Oct 00:39 → 05 Oct 20:29 · sha256 0eb44f139f93

Solana Launch Dataset v16

8,939 snapshots

1,794 launches · 05 Oct 00:39 → 05 Oct 19:06 · sha256 07b5c66cb3c8

Solana Launch Dataset v15

9,754 snapshots

1,957 launches · 05 Oct 00:39 → 05 Oct 19:03 · sha256 bbc271ff0196

Solana Launch Dataset v14

9,235 snapshots

1,853 launches · 05 Oct 00:39 → 05 Oct 17:47 · sha256 83676a982760

Solana Launch Dataset v13

2,018 snapshots

406 launches · 05 Oct 00:39 → 05 Oct 15:47 · sha256 1cbbdf250876

Solana Launch Dataset v12

7,235 snapshots

1,452 launches · 05 Oct 00:39 → 05 Oct 13:45 · sha256 d45a86d8a29b

Solana Launch Dataset v11

7,225 snapshots

1,450 launches · 05 Oct 00:39 → 05 Oct 13:42 · sha256 1a237aeb2f86

Solana Launch Dataset v10

7,170 snapshots

1,439 launches · 05 Oct 00:39 → 05 Oct 13:36 · sha256 94c2078b7766

Solana Launch Dataset v9

3,937 snapshots

789 launches · 05 Oct 00:39 → 05 Oct 07:16 · sha256 78a0a22f0364

Solana Launch Dataset v8

3,757 snapshots

753 launches · 05 Oct 00:39 → 05 Oct 07:01 · sha256 7a3cc82a36d5

Solana Launch Dataset v7

3,122 snapshots

626 launches · 05 Oct 00:39 → 05 Oct 06:02 · sha256 8f5197a2a59b

Solana Launch Dataset v6

2,492 snapshots

500 launches · 05 Oct 00:39 → 05 Oct 05:00 · sha256 35c7f3706ceb

Solana Launch Dataset v5

784 snapshots

157 launches · 05 Oct 00:39 → 05 Oct 02:24 · sha256 a8050102d8fd

Solana Launch Dataset v4

1,170 snapshots

235 launches · 05 Oct 00:39 → 05 Oct 03:01 · sha256 ebbadfe5c81d

Solana Launch Dataset v3

694 snapshots

139 launches · 05 Oct 00:39 → 05 Oct 02:12 · sha256 7e44dbd62410

Solana Launch Dataset v2

115 snapshots

23 launches · 05 Oct 00:39 → 05 Oct 01:12 · sha256 30708f586621

Solana Launch Dataset v1

60 snapshots

12 launches · 05 Oct 00:39 → 05 Oct 00:44 · sha256 010016f628b3