Skip to content

Training data for trading models.

Screens, voices, fills and labels from real Solana traders, joined per session. The corpus is collecting and every count here is live.

Approved sessions…
Approved hours…
Narrated sessions…
Verified fills…
Contributors…

Screen sessions

…Collecting

Full-screen recordings of real traders working real terminals, redacted in the browser before upload.

Continuous screen captures of discretionary trading on Solana venues: StonkFun, pump.fun, Axiom, Photon, Jupiter and xStocks. Redaction boxes are burned into the pixels client-side, so balances, seed phrases and DMs never leave the trader's machine. Every session is watched by a person before it enters the corpus.

Annotations
Review scoreRedaction countVenueSession window
Venues on tape
StonkFunpump.funAxiomPhotonJupiterxStocksGMGNBullXRaydiumMeteora
Built for
Computer-use agents for trading UIs, Screen-to-action models, UI state understanding
FieldType
session_idStable id for the recordinguuid
contributorPseudonymous contributor id (hashed wallet)string
videoWebM or MP4, original resolution, redactions burned inuri
duration_secLength of the recordingnumber
started_atWall-clock start, used to align fillstimestamp
redactionsNumber of redaction boxes appliedinteger
review_scoreHuman reviewer quality scoreinteger 1-5

Narrated sessions

…Collecting

Screen sessions with the trader's own voice explaining what they see and why they act.

The subset of screen sessions with a mic track. Traders think out loud: thesis, levels, sizing, doubt, the exit. Pairs naturally with the wallet tape so every spoken decision has a timestamped on-chain fill next to it.

Annotations
TranscriptSpeech segmentsLanguage
Venues on tape
StonkFunpump.funAxiomPhotonJupiterxStocksGMGNBullXRaydiumMeteora
Built for
Reasoning traces for trading agents, Vision-language alignment, Intent prediction
FieldType
session_idJoins to screen sessionsuuid
audioNarration track, separate from screen audiouri
transcriptWord-level timestamps (delivered on license)json
languageBCP 47 language tagstring
narration_secSeconds of speech in the sessionnumber

Wallet tape

…Collecting

Verified on-chain executions from consenting traders, aligned to the second with their screen.

Swaps pulled from the trader's linked wallet after they sign a consent message. Each execution is normalized (side, mint, size, SOL and USD value, venue, signature) and attached to the session whose time window contains it. Ground truth for what the screen was leading up to.

Annotations
SideVenueSession alignment
Venues on tape
StonkFunpump.funAxiomPhotonJupiterxStocksGMGNBullXRaydiumMeteora
Built for
Action labels for imitation learning, Execution quality studies, Outcome modeling
FieldType
signatureSolana transaction signaturestring
timeBlock timetimestamp
sideDirection relative to the quote assetbuy | sell | swap
mintTraded token mintstring
amountToken amount, UI unitsnumber
sol_amountSOL leg when the quote is SOLnumber
usd_valueUSD value at execution when knownnumber
sourceVenue or aggregator reported by the indexerstring
session_idSession the fill landed inuuid | null

Decision labels

…Collecting

Trader-written labels for each session: setup, venue, assets, outcome and a short note.

Optional labels the trader adds before submitting: tags (entry, exit, scalp, swing, snipe, DCA, cut loss), venues, assets, outcome and a short note on what they were thinking. Losses are labeled as losses.

Annotations
Setup tagsOutcomeFree-text rationale
Venues on tape
StonkFunpump.funAxiomPhotonJupiterxStocksGMGNBullXRaydiumMeteora
Built for
Reward modeling, Strategy classification, Post-trade review agents
FieldType
session_idJoins to screen sessionsuuid
tagsSetup tags from a fixed vocabularystring[]
venuesVenues used in the sessionstring[]
assetsTickers or mints tradedstring[]
outcomeSelf-reported resultwin | loss | flat | null
noteWhat the trader was thinking, in their wordsstring