Hugging Face's logo

Datasets:
Slinky21
/
Pumpfun_Memecoin_Corpus

Tasks:
Tabular Classification
Tabular Regression
Time Series Forecasting
Modalities:
Text
Formats:
parquet
Size:
10M - 100M
Tags:
solana
cryptocurrency
memecoin
defi
pump.fun
blockchain
Libraries:
Datasets
Dask
Polars
License:
Dataset card Data Studio Files
xet
 Community
3
Pumpfun_Memecoin_Corpus / KNOWN_ISSUES.md
Slinky21's picture
Slinky21
Update KNOWN_ISSUES.md
56b0ed9 verified
preview code
|
Raw History Blame Contribute Delete
15 kB

Known Issues — PumpFun Launch-to-Graduation Corpus

This document is the authoritative, complete record of every data-quality issue found in this corpus, how it was confirmed, its exact scale, and how to handle it. Every number below was independently verified by direct querying against the real data — nothing here is assumed or inherited without a check.

Read this before running any analysis. Several items below are not optional to handle correctly — skipping them will produce silently wrong results, not an error.

1. Fixed issues (already corrected in tokens.parquet)

1.1 Mayhem-mode 2B-supply concentration bug

Mayhem Mode tokens mint 2,000,000,000 tokens, double the standard 1,000,000,000 — but concentration percentages (initial_top1/5/10_pct, dev_buy_pct) were originally computed against a hardcoded 1B-supply assumption for every token, mayhem or not. Proof of the bug: pre-fix, initial_top10_pct reached values up to 200% for mayhem tokens — a mathematically impossible percentage.

  • Scale: 109,561 tokens (~12.7% of the corpus)
  • Fix applied: *_corrected columns (initial_top1_pct_corrected, initial_top5_pct_corrected, initial_top10_pct_corrected, dev_buy_pct_corrected, supply_bug_corrected) — use these, not the originals, for any concentration analysis.

2. Unfixable, excluded issues

2.1 Bonding-curve-balance leak

A separate, smaller set of tokens show impossible concentration values (median top1_pct 98.8%, up to 129.3% at top10) caused by a check meant to exclude the bonding curve's own token balance from the "largest holder" calculation failing to fire. Confirmed via correlation: top1_pct correlates 0.71 with curve balance at first trade for affected tokens, vs. -0.02 for unaffected tokens.

  • Scale: 1,486 tokens, concentrated in a Jun 23-24 window tied to a burst of clone/factory-launch creator activity
  • Handling: No correct value can be recovered — these rows are flagged via top10_pct_suspect, and should be excluded, not corrected or imputed.

3. Trade-level data issues (trades.parquet)

3.1 sol_amount internally inconsistent with token_amount x price_sol

For a genuine subset of trade rows, sol_amount does not match token_amount x price_sol — in the worst cases, off by up to ~39,000,000x. Root cause not fully diagnosed (a units/decimals mismatch somewhere in ingestion is a plausible candidate, not confirmed).

  • Scale: 1,136,321 rows (3.38% of 33,581,765)
  • Detection: flag any row where sol_amount / (token_amount * price_sol) falls outside [0.01, 100]
  • Handling: exclude from any SOL-denominated computation (volume, P&L, leaderboards). This filter is large enough to fully distort a naive top-wallet ranking if skipped — confirmed during this project's own research.

3.2 sol_amount / price_sol NULL (missing, not corrupted — distinct issue)

Separately from §3.1, a real share of trades have sol_amount and price_sol as NULL rather than wrong — the price was simply never captured for that trade.

  • Scale: 2,360,061 rows (7.03%)
  • Pattern, confirmed via direct investigation:
    • Heavily front-loaded in time — the corpus's first ~10 days carry a disproportionate share, peaking at 33.6%/45.3% null on Jun 9-10 (the same 2-day window as the holder-count break, §5.1 — an interesting coincidence, not proven to share a root cause).
    • Not tied to trade direction (buy 7.3% vs. sell 6.7% — similar) or to a specific ingestion source.
    • A whole-token phenomenon, not scattered per-trade gaps: of ~20,300 affected mints, over 99.8% are tokens where every trade is null-price, not a handful of trades within an otherwise-fine token.
    • Skews toward very early trades in a token's life (median ~20 min post-launch among null rows, vs. ~2.4h corpus-wide).
  • Handling: flagged via sol_price_missing — kept deliberately separate from sol_amount_suspect (§3.1) so the two issues are never conflated.

3.3 Non-human wallet contamination

The address BwWK17cbHxwWBKZkUYvzxLcNQ1YVyaFezduWbtm2de6s is the Solana System Program, confirmed via block explorer (Owner: System Program, isOnCurve: FALSE) — not a market participant.

  • Scale: 6,690,901 trade rows (19.92% of all trades), touching ~35% of all mints in the corpus
  • Handling: exclude this address from any wallet-derived feature — unique wallet counts, wallet concentration, P&L leaderboards, wash -trading signals. Left in, it silently inflates unique-wallet counts by one and contaminates every wallet-concentration metric on a third of the corpus.

3.4 curve_pct_depleted / curve_pct_depleted_eob values outside [0,100]%

Both the per-trade (trades.curve_pct_depleted) and end-of-bucket (snapshots.curve_pct_depleted_eob) versions of this field show values outside the theoretically valid [0,100]% range.

  • Scale: 1,215,810 trade rows (3.62%), 730,670 snapshot rows (2.71%)
  • Magnitude, confirmed via direct distribution check: all bad values are capped at exactly 110.0, with 99.95%+ of them barely over the 100% boundary (in the 100-200 bucket, essentially none above 200) — this looks like a bounded, designed overshoot mechanism, not open-ended corruption.
  • Mayhem-mode correlation: real but weak (5.33% bad for mayhem tokens vs. 3.26% non-mayhem, ~1.6x) — too weak to be the primary mechanism; the known supply bug (§1.1) produces a much starker gap when it's the actual cause, confirmed via a synthetic test during this project's own research.
  • Temporal pattern (the dominant signal): flat and low (0.02%-1.7%) for the entire corpus from Jun 5 through Jul 2, then jumps sharply to 8-11% starting Jul 4 and holds through the corpus end (Jul 14).
  • Jul 3 is a confirmed one-day gap — zero recorded trades that day, due to a websocket collection outage. This explains the missing day itself. It does not, by itself, fully explain why the badness rate shifted upward and stayed elevated for the ten days following recovery — whether that reflects a config/version change made during outage recovery, a platform-side change coinciding with the same window, or something else has not been further diagnosed.
  • Handling: treat as a fifth, distinct regime break (§5.5). Any model or analysis using curve-derived features (curve_velocity, curve_accel, curve_pct_depleted_at_horizon) should account for this shift, especially for the final third of the collection window.

4. wallet_stats.parquet reliability — read this before using this table at all

This table's own activity/volume columns cannot be trusted without independent verification.

  • Systematic cross-check (100,000 sampled wallets against a live recount from trades.parquet): only 37.7% exact match on total_trades (59.7% exact match on tokens_traded). Mean ratio of (real / reported) trades = 15.9x, median = 2x, max = 16,055x.
  • Root cause, confirmed via a targeted diagnostic distinguishing three competing hypotheses: a stale snapshot, not a truncated aggregation window or a random sampling bug. The discrepancy ratio climbs monotonically from ~3.3x (wallets whose last real trade fell in the corpus's first week) to ~46.2x (wallets whose last real trade fell in the final week) — the exact signature of a table built once, early or mid-collection, and never regenerated.
  • What is affected: tokens_traded, total_trades, total_buy_volume_sol, total_sell_volume_sol. Do not use these four columns for any activity-volume analysis — recompute directly from trades.parquet instead (see quickstart.py for the pattern).
  • What is not (confirmed) affected: first_seen_at/last_seen_at timestamps were not part of the cross-check that found this issue — they may still be usable, though they were not independently verified either. Treat with appropriate caution, not blind trust .

4.5. outcome_label has no 'rug' value — using it as one will silently

return zero rows. Real values: dead, pump_dump, slow_bleed, major_pump, minor_pump, sustained. quickstart.py's Section 5 (Gini-vs-rug metric) filters on outcome_label = 'rug' and is currently broken as a result. Use rug_detected (boolean, in postgard_outcomes.parquet) for a direct rug flag, or group dead/pump_dump/slow_bleed as "failed" outcomes if you want the broader category. Fix pending.

4.55. migrations.parquet.pool_address contains synthetic sentinel values,

not real pool addresses, for ~25% of rows. synthetic_graduation_queue (1,380 rows) marks graduations queued but not yet matched to a real pool; backfilled_from_pumpswap_trade (69 rows) marks pool addresses reconstructed after the fact from PumpSwap trade data rather than observed directly at migration time. Filter these out if you need a real on-chain pool address, or check pool_address NOT IN ('synthetic_graduation_queue', 'backfilled_from_pumpswap_trade') before joining downstream.

Items 4.5 and 4.55 reported by M. Huzaifa (July 2026) — thank you.

5. Regime breaks / non-stationarity

Five separate, independently-confirmed step-changes exist within the same 39-day corpus. Any cross-corpus time comparison, or any train/test split whose boundary straddles one of these dates, should account for it.

5.1 Jun 9-10: initial_holder_count hard-cap change

Maximum value changed from exactly 54 to exactly 20 across a 2-day transition, and never reverted. initial_top1/5/10_pct move in lockstep, almost certainly because holder count is the denominator context for those percentiles. Ruled out as a collection outage — hourly trade volume across Jun 5-20 shows no gap at this boundary. Working hypothesis, not independently confirmed: Solana's getTokenLargestAccounts RPC method is natively capped at 20 accounts by protocol design, and this date marks a change in (or loss of a workaround for) that native limit. Left unfixed in the data — this is a real, permanent feature of the corpus, not something to patch.

5.2 Jun 20-22: Mayhem Mode rollout

A new token type (2B supply, see §1.1) did not exist before this date and ramped to ~28-30% of new launches within three weeks. A real product change, not a bug — but means token-type composition is not stationary across the collection window.

5.3 Jun 23-24: Second concentration-bug window

Overlaps with the bonding-curve-balance leak (§2.1) and a burst of clone-launch creator activity in the same window.

5.4 Jun 24-25: Platform-wide silence-rate drop

The share of tokens with zero post-launch trading activity dropped from ~96.4% to ~88.9%. Checked and ruled out as a polling-cadence artifact (sampling interval unchanged across the boundary) — most likely reflects genuinely rising launch volume on the platform.

5.5 Jul 3-4: Websocket outage + curve_pct_depleted shift

See §3.4 for full detail. Jul 3 is a confirmed collection outage (zero trades that day). The subsequent shift in curve_pct_depleted badness rate (flat <2% before, 8-11% after, sustained through corpus end) is temporally associated with this outage but not fully root-caused beyond that.

6. Schema & naming gotchas

  • outcome_label = 'dead' (a token that graduated, then died within 24h) is a completely different population from a token that never graduated at all. Same word, two unrelated meanings across different parts of this corpus — do not conflate them.
  • postgard_outcomes.parquet's real exported schema does not match older collector source-code comments. Fields like label_time_top10_pct / label_time_gini are defined in source-level DDL but do not exist in the actual exported file. Always verify against a live DESCRIBE, never trust schema from code comments or documentation.
  • postgard_snapshots.parquet covers more distinct mints (5,701) than a typical filtered target population (e.g. graduated, non-mayhem tokens is a smaller subset) — filter explicitly by your actual target population; don't assume table membership implies population membership.
  • v_tokens_bonding_curve / v_sol_bonding_curve (in trades.parquet) are raw micro-token units, not human-scale token counts. A naive /1e9 conversion produces values off by orders of magnitude (though correlations computed on the wrongly-scaled values remain valid, since it's a consistent scaling error). The correct relationship is price_sol = (v_sol_bonding_curve / v_tokens_bonding_curve) / 1000.

7. Modeling gotchas (not data bugs, but easy to leak on)

  • entry_price_20s_usd / entry_price_30s_usd / entry_price_1m_usd are a near-perfect label proxy if used naively as predictive features. Null rate is ~0.14-0.18% for graduated tokens vs. ~93.5-93.6% for non-graduated tokens, with no time-to-graduation gradient — these fields are populated almost exclusively because a token graduated, not because of when. Suppressing the missingness indicator column is not sufficient to prevent leakage — tree-based models (XGBoost, LightGBM) learn native NaN split directions and will still effectively split on "is this field present." If building any graduation-prediction model on this corpus, exclude these three fields entirely rather than indicator-suppress them.
  • first_trade_price_sol was checked with the same methodology and is fine — its null gap is small and roughly even across outcomes (27.9% graduated vs. 25.5% non-graduated), not the same near-total separator as the fields above.

8. What's confirmed clean (checked, not just assumed)

  • market_cap_sol / price_sol in trades.parquet are correctly supply-invariant (derived from raw bonding-curve reserve ratios, not a percentage-of-total-supply calculation) — confirmed clean of both the mayhem-mode bug (§1.1) and the bonding-curve-balance leak (§2.1). Don't assume every supply-related field is contaminated just because nearby fields are — each was checked individually against its actual formula.

9. Tooling note (DuckDB users)

If querying this corpus with DuckDB, USING SAMPLE N (reservoir method) has been observed to silently return 0 rows on filtered queries in some environments, even though the identical query without the sample clause returns thousands of rows — not root-caused against DuckDB internals, but reproducible via direct A/B testing. USING SAMPLE N (bernoulli method) throws an outright parser error when combined with a discrete row-count target. Workaround that was verified to work correctly, including with joins, at matching scale: use setseed() + ORDER BY random() + LIMIT instead of USING SAMPLE entirely.