jrhode666/gpt-scorer

GPT Scorer — Pump.fun Token Migration Predictor

XGBoost model trained on 135K+ pump.fun token launches to predict which tokens will migrate to a DEX (PumpSwap/Raydium/Meteora/Jupiter) vs rug/stall/abandon.

Performance

Metric	Value
ROC AUC	0.9899
Precision @ 80%	80.2%
Recall @ 80%	78.3%
F1 (migrated)	0.66
Training samples	135,617
Features	36
Trained	Daily at 6am UTC

Top Features

1. entity_total_migrations — how many times entity has migrated before
2. entity_migration_rate — entity success rate
3. inactivity_gap_seconds — dead time between trades
4. is_quality_dev — archetype classification
5. distinct_sellers
6. total_sells
7. is_bot_cluster — bot detection
8. curve_pct_filled — bonding curve progress
9. entity_prior — Bayesian prior score
10. distinct_buyers

Usage

import xgboost as xgb
from huggingface_hub import hf_hub_download

# Download model
model_path = hf_hub_download(repo_id="jrhode666/gpt-scorer", filename="xgb_scorer.json")

# Load
model = xgb.XGBClassifier()
model.load_model(model_path)

# Predict (36 features required — see xgb_meta.json)
probability = model.predict_proba(features_df)[:, 1]

Features (36 total)

See xgb_meta.json for the full feature list. Core categories:

• Activity: unique_buyers_15s/30s/60s, buy_count_*, sell_count_*, total_buys, total_sells
• Microstructure: early_sell_pressure, buyer_concentration, repeat_buyer_rate
• Entity history: entity_prior, entity_confidence, entity_migration_rate, entity_rug_rate, entity_total_launches, entity_total_migrations
• Archetype: is_quality_dev, is_moderate_dev, is_whale_dev, is_serial_rugger, is_bot_cluster
• Temporal: launch_hour_utc, inactivity_gap_seconds
• Market: curve_pct_filled
• Derived: buy_sell_ratio, seller_buyer_ratio, buyers_per_minute

Target

Binary classification: 1 if token migrated to DEX (PumpSwap, Raydium, Meteora, Jupiter), 0 otherwise.

License

MIT