bryceweiner/fly-trader-seed

fly-trader training database

The PostgreSQL database fly-trader trains on: every PumpSwap trade of tokens that graduated from pump.fun, aggregated to 1-minute candles with order-flow, wallet and launch facts, from 2026-04-19 to 2026-09-21. This is what the shipped model was trained from; restore it to retrain on master without re-downloading five months of archive.

Not included: runtime state (positions, decisions, the fly's learning tables), API logs, the ephemeral swap tapes, and the per-day wallet-skill tables (30 GB of parquet; master rebuilds them from these tables with fly-trader fit-wallet-skill).

Contents

fly_trader_core.dump — a pg_dump -Fc (custom format, compressed) of these tables:

Table	What
pump_minutes	1-minute candles per token: OHLC, buy/sell SOL, trade counts, buyers, wash, top-holder and insider selling, buys by wallet-skill decile
pump_events, pump_pools	pool creations and graduations; PumpSwap pools with their reserves
tokens, token_stats, token_insiders	token metadata and decimals; Jupiter token stats history; creator, bundle and early wallets per token
corpus_meta, corpus_tokens, corpus_features	launch facts per token (bundle share, dev holdings, graduation top-10 share, curve state) and the training feature manifest
mature_days, replay_days, replay_hours	which archive days and hours are complete
schema_version	6

Restore

createdb fly_trader
pg_restore -d fly_trader --no-owner fly_trader_core.dump
git clone https://github.com/bryceweiner/fly-trader.git && cd fly-trader && uv sync
uv run fly-trader init-db          # adds the runtime tables the dump leaves out

Then uv run fly-trader build-mature / train-selector as described in the repository README. PostgreSQL 15 or newer.

Provenance and licence

Trades from the PumpAPI archive and stream; token stats from Jupiter. Public on-chain data, aggregated. MIT, no warranty. A ready-to-run trader built from this data is at bryceweiner/fly-trader.