Project planning notes:
docs/PLAN.md.
Language / 语言: English primary · 中文概览如下。
用于训练和评测 Cultivating ML Agent 的公开分类基准项目，记录实验、失败分析和能力沉淀。
A public classification benchmark used to train and evaluate TopPrism's Cultivating ML Agent.
LEARNING PROJECT · TABULAR ML · KAGGLE BENCHMARK
This is not a customer product. It is a measurable environment for experimentation, failure analysis, ensemble design, and knowledge crystallization.
Three-class irrigation-need classification with Balanced Accuracy.
The repository records a progression from a LightGBM baseline through target encoding, pseudo-labeling, diverse ensembles, and external prediction integration.
The GitHub About text currently says Best LB = 0.97785, while the README reports later results up to 0.98150.
Unify:
- README;
- GitHub About;
- badges;
- any project index.
Recommended About:
Learning project for Cultivating ML Agent --- Kaggle PS S6E4 tabular classification; experiment history through best reported Public LB 0.98150.
The README already contains a more valuable story than the leaderboard number:
- pairwise target encoding produced meaningful gains;
- iterative pseudo-labeling can degrade performance;
- ensemble source quality matters more than source count;
- self-trained models plateaued before external-prediction integration;
- diverse signals can matter more than adding more similar models.
These lessons are reported first on this page.
The full R01--R18 experiment log is preserved verbatim in
docs/competition-log.md as project history.
This README records only the headline progression:
Baseline (R01 LightGBM)
-> best self-trained (R09 10-model + pseudo + stacking, LB 0.97785)
-> key failed experiment (R12 iterative pseudo-labeling hurt performance)
-> best externally-assisted (R17 Schema8 + formula prediction, LB 0.98150)
-> reusable skills crystallized for the Cultivating ML Agent
Because later results integrate external prediction sources, clearly separate:
- self-trained performance;
- externally assisted / blended performance.
Do not imply the final LB is produced solely by the repository's own trained models.
topprism:
purpose: learning-project
capability: tabular-ml
maturity: learning
evidence:
type: kaggle-benchmark
best_reported_public_lb: 0.98150
parent:
- cultivating-ml-agent