poker44-my-model (top_blend)
Original bot-detection model for Poker44 (Bittensor subnet 126).
Model description
top_blend is a weighted-average ensemble (60/40) of two independently
trained variants:
- rank1_gbm — LightGBM gradient-boosted classifier over chunk-level behavioral features
- rank3_stacked — out-of-fold stacked ensemble (ExtraTrees + RandomForest
- HistGradientBoosting + LightGBM) with a logistic-regression meta-learner
Both members apply isotonic calibration and a conformal false-positive-rate safety shift on top of the raw model score, matching the validator's threshold-based scoring at 0.5.
Training data
Public Poker44 training benchmark (https://api.poker44.net/api/v1/benchmark),
15 release dates (2026-07-01 through 2026-07-15). No private or
validator-only evaluation data was used.
Evaluation
Leave-one-date-out cross-validation using the validator's exact reward()
formula (poker44/score/scoring.py in the
Poker44-subnet repo):
| Variant | Calibrated OOF reward | AP | Bot recall @5% FPR | Hard FPR |
|---|---|---|---|---|
| top_blend | 0.9216 | 0.9500 | 0.7971 | 0.0000 |
| rank1_gbm | 0.9201 | 0.9480 | 0.7945 | 0.0000 |
| rank2_drse | 0.9148 | 0.9486 | 0.7760 | 0.0000 |
| rank3_stacked | 0.9125 | 0.9500 | 0.7668 | 0.0000 |
Training code
Full training pipeline (data download, feature engineering, model definitions, calibration, cross-validation): see the source repository.
Intended use
Miner inference for Poker44 subnet 126 DetectionSynapse chunks — one
bot-risk score in [0, 1] per chunk of poker hands.
Attestation
- Trained only on the public benchmark endpoint above.
- No validator-only or private data used.
- Open source, MIT licensed.