Posts

TabICL exposes a built-in fine-tuning pipeline via FinetunedTabICLClassifier. On five real-world classification datasets, I compared zero-shot TabICL against fine-tuned TabICL (30 epochs, early stopping, validation-driven hyperparameter selection). The result: fine-tuning helps on some datasets, hurts on others, and never moves AUC by more than ±0.7 pp. On telco-churn it is consistently beneficial (+0.16 to +0.59 pp). On cc-fraud it is completely flat — zero-shot is already near-perfect. The only consistent signal is that fine-tuning with too little data or the wrong seed can degrade performance.

TabICL exposes a built-in fine-tuning pipeline via FinetunedTabICLClassifier. On five real-world classification datasets, I compared zero-shot TabICL against fine-tuned TabICL (30 epochs, early stopping, validation-driven hyperparameter selection). The result: fine-tuning helps on some datasets, hurts on others, and never moves AUC by more than ±0.7 pp. On telco-churn it is consistently beneficial (+0.16 to +0.59 pp). On cc-fraud it is completely flat — zero-shot is already near-perfect. The only consistent signal is that fine-tuning with too little data or the wrong seed can degrade performance.

Hermes Agent orchestrates two persistent machines — a free-tier ARM64 VPS and a custom x86-64 workstation — to run Rust and PyTorch workloads without sandbox churn.

Hermes Agent orchestrates two persistent machines — a free-tier ARM64 VPS and a custom x86-64 workstation — to run Rust and PyTorch workloads without sandbox churn.

Stacking TabPFN3, TabICL, and XGBoost provides at most +0.5 pp AUC on most tabular datasets. But on heavily imbalanced fraud detection, the ensemble is dramatically more robust. The reason is not model diversity in the abstract—it is concrete feature disagreement. XGBoost and TabPFN disagree strongly on which features matter for fraud (Spearman ρ = 0.24), while they agree closely on every other dataset (ρ = 0.67–0.95). When models look at different features, stacking hedges correlated failure modes. When they look at the same features, stacking is just expensive averaging.

Stacking TabPFN3, TabICL, and XGBoost provides at most +0.5 pp AUC on most tabular datasets. But on heavily imbalanced fraud detection, the ensemble is dramatically more robust. The reason is not model diversity in the abstract—it is concrete feature disagreement. XGBoost and TabPFN disagree strongly on which features matter for fraud (Spearman ρ = 0.24), while they agree closely on every other dataset (ρ = 0.67–0.95). When models look at different features, stacking hedges correlated failure modes. When they look at the same features, stacking is just expensive averaging.

PFN wins below 10k rows, ICL catches up by 100k, and PFN degrades beyond 200k. Both are 2× apart in speed because PFN is 2× larger. We also found a clean 21% inference speedup with bfloat16 autocast.

PFN wins below 10k rows, ICL catches up by 100k, and PFN degrades beyond 200k. Both are 2× apart in speed because PFN is 2× larger. We also found a clean 21% inference speedup with bfloat16 autocast.

Two #[inline] annotations on the innermost recv path improve large-object throughput by 14.7% and medium objects by 11% with no regressions.

Two #[inline] annotations on the innermost recv path improve large-object throughput by 14.7% and medium objects by 11% with no regressions.