Datasets:

---

shaypal5

/

leadforge-lead-scoring-v1

like 1

Hi everyone 👋

I'm Shay, the author of this dataset. Welcome — and thank you for being here.

LeadForge is a reproducible, three-tier synthetic B2B sales funnel dataset for teaching lead scoring. All three tiers — Intro, Intermediate, and Advanced — share the same relational schema and the same task (converted_within_90_days), varying only in conversion rate, noise, and missingness.

🔨 The challenge: try to break it

This dataset ships with a break-me guide that catalogues nine adversarial patterns to look for. Here are the ones I'm most curious whether you find:

1. The leakage trap — total_touches_all is a deliberately leaky column. Can you detect it programmatically before reading the docs?
2. Split contamination — ≈93% of test accounts also appear in train. What's the AUC gap between a random split and a proper GroupKFold(account_id) split?
3. Calibration drift — the Advanced tier has a median calibration max-bin error of 0.221 with very high seed-to-seed variance. Can you fix it?
4. GBM not beating LR — on all three tiers, gradient boosting slightly underperforms logistic regression on the raw snapshot. Can you figure out why, and fix it with feature engineering?
5. Channel signal — lead_source has near-random univariate AUC (~0.50–0.52). Is that realistic? Is there a way to extract signal from it?

💬 Feedback I'd love

• Does the difficulty feel right for each tier?
• Does the B2B SaaS scenario feel realistic? What would make it more so?
• Missing a feature you'd expect in a real CRM? Let us know.
• Found something that looks like a bug? File it at leadforge-dev/leadforge.

The five starter notebooks on the Code tab (Starter, EDA, Feature Engineering, Calibration, SHAP) are the quickest way to get oriented.

Looking forward to seeing what you find!
— Shay