Ashxr/solana-memecoin-lifecycle-v0

solana-memecoin-lifecycle-v0

Pump.fun token launch + lifecycle trade data pulled from Dune's dex_solana.trades.

Files (uploaded manually via HF web UI — see below)

• —candidates_step1_500mints.csv — 500 new pump.fun launches, 2025-11-15 00:00:01–00:43:11 UTC (single cohort, one dense 43-minute burst)
• —01M2BB9QVSNF2V2JWDPE4TWRCK.csv — full 68,618-row trade history for those 500 mints, 48h window
• —cohort_A.parquet — 186 mints, launched 2025-11-08 08:00–08:30 UTC, 18,559 trades, 48h-truncated
• —cohort_B.parquet — 357 mints, launched 2025-11-15 14:00–14:30 UTC, 43,593 trades, 48h-truncated
• —cohort_C.parquet — 397 mints, launched 2025-11-22 20:00–20:30 UTC, 99,293 trades, 48h-truncated
• —cohort_D.parquet — 324 mints, launched 2025-11-29 03:00–03:30 UTC, 21,478 trades, 48h-truncated
• —cohorts_ABCD_first_seen.parquet — first-trade timestamp per mint across cohorts A–D

Total: 1,764 unique tokens across 5 independent launch windows (different days/times of day), zero mint overlap between the original 500 and cohorts A–D.

Known caveats (read before modeling)

• —Zero Raydium trades across every cohort (5/5) — post-graduation activity for these tokens appears to stay entirely on pumpswap, not migrate to Raydium. Consistent across 5 independent samples, treated as a real pattern, not a query bug.
• —Cohorts A–D were originally pulled with a shared, non-per-mint time filter that let some cohorts' trade history run up to ~22 days past launch instead of 48h. This has been corrected: all rows in the uploaded cohort_*.parquet files are truncated to each mint's own first_trade_time + 48h. The raw (untruncated) CSV export is not included here.
• —`log_messages`-style narrative fields do not exist in this schema at all (this is Dune's decoded trade table, not raw instructions) — labels/features must be built from token_bought_amount/token_sold_amount/amount_usd deltas over time, not from log parsing.
• —Trades-per-mint is heavily right-skewed in every cohort (many 1–2 trade "stillborn" tokens, a few outliers with 10,000+ trades) — expect this in any label distribution.