In this article, you will learn how to compare two strategies for adapting a large language model to a specific domain: fine-tuning with LoRA versus...

In this article, you will learn how to compare two strategies for adapting a large language model to a specific domain: fine-tuning with LoRA versus...

In this article, you will learn how to compare two strategies for adapting a large language model to a specific domain: fine-tuning with LoRA versus...

4Paradigm and Beihang University won the KDD Cup 2020 Reinforcement Learning Track, while Chinese teams took the top three places in the AutoML Track that 4Paradigm organised with ChaLearn, Stanford University and Google.

Available on github https://github.com/hariomlohardev/mini-sklearn KMeans matched sklearn to the decimal while every classifier trailed. ## What I built `minisklearn`: five from-scratch models plus hand-rolled metrics, all stdlib Python, trained on the Titanic dataset (889 rows, 7 features, ~38% survived). Same split throughout (`test_ratio=0.2, seed=0`), all scored with my own confusion-matrix metrics. ## The surprise All the other models that i tested against the sklearn library are performing a bit bad than sklearn implementation but the KMeans matched exectly each things no differance I didn't expect that to happend ## The numbers Same split, same ruler — my models vs `scikit-learn`: model | mine (acc / P / R / F1) | sklearn (acc / P / R / F1) | gap | verdict ---|---|---|---|--- LogisticRegression | 0.684 / 0.714 / 0.294 / 0.417 | 0.847 / 0.806 / 0.794 / 0.800 | 0.164 | investigate DecisionTree | 0.706 / 0.682 / 0.441 / 0.536 | 0.780 / 0.716 / 0.706 / 0.711 | 0.073 | investigate RandomForest | 0.695 / 0.652 / 0.441 / 0.526 | 0.831 / 0.828 / 0.706 / 0.762 | 0.136 | investigate KMeans | inertia 742754.6 | inertia 742754.6 | 0.0 | match The pattern: my classifiers systematically under-predict "survived" (recall 0.29–0.44 vs sklearn's 0.71–0.79) A too-tiny learning rate on unscaled features plus exact ID3 splits fragmenting continuous inputs. KMeans, meanwhile, matches sklearn's inertia to the decimal, which says the Lloyd loop is exactly right. ## What I'd do differently In future i will follow this models approch if something can be solved by this rather than jumping to Neural networks and gradient decent I know there are problems that can't be solved by this tiny models but they surpries so much well on some special cases where neutral networks perform bad

Image of Reshama Shaikh (a dark-haired person wearing an blue colored shirt). The image has a darker background with nebula and puzzle pieces matching together.

Image of Reshama Shaikh (a dark-haired person wearing an blue colored shirt). The image has a darker background with nebula and puzzle pieces matching together.

Findings [1] 2026-09-22 🔬 An Oscar, Two Asteroids, and the Algorithm in Your sklearn: John Platt on AI for Science How often do you get to talk to a guest…

flannrust is a bit-exact Rust port of nanoflann, the C++ kd-tree library. Faster parallel builds, identical search results, and Python bindings that outperform SciPy's cKDTree. Benchmarked end to end.

sklearn-compatible PyTorch implementations of Soft Decision Trees, HMoE, and classifier comparison tests (5×2cv F-test). pip install neural-trees - cgrtml/neural-trees