README.md

metadata

title: Corvana Bank LimitWise
emoji: 💳
colorFrom: green
colorTo: gray
sdk: gradio
sdk_version: 6.31.0
python_version: '3.12'
app_file: app.py
pinned: false
license: cc-by-4.0
models:
  - drdavidprivacy/limitwise-model
datasets:
  - drdavidprivacy/limitwise-data
  - drdavidprivacy/practicum-case-packs
short_description: LimitWise v2, the signed model and its repositories

LimitWise v2: Corvana Bank

Kept by: Tom Becker, AI Governance Lead. Signed by: Anika Rao, Chief AI Officer (Week 6 memo).

Program: AI Governance Practicum, Track 1, cohort 1

This page is the public home of a governed system. The model is LimitWise v2, the version the Chief AI Officer signed in Week 6. The app has four tabs: one account (the three Week 6 cases and the record's own reasons), a batch upload (score and rank a file), monitoring (read a month against the reference file with Evidently's data-drift report), and the repositories (the links below).

The signed version

• File: limitwise_v2.pkl, 18,636 bytes, SHA-256 7842edc29acd316873ffa482db705ab2ea09427f8b4e1186436713d316de8ecb
• Recalibration object: limitwise_v2_recalibrated.pkl, 89,015 bytes, SHA-256 b6508dfcc69ca108e3f08fc2be0843dab1d2517355970b8e622e52a34127ca37 (the isotonic recalibration the Week 6 memo ratified; the odds correction to the 3 percent book rate, 0.1089, is applied in the app)
• Plain-language reasons: reasons.csv (one reason for every value of every input)
• Operating point: the top 10 percent of the ranked file
• Libraries: scikit-learn 1.9.1, fairlearn 0.15.0
• Data: the remediated working file in the program's dataset repository, loaded by link, never copied here: https://huggingface.co/datasets/drdavidprivacy/practicum-case-packs/resolve/main/corvana_modeling_extract_remediated.csv (MD5 f79f13a5119f78572e9f6cc3243ea6ae)

Monitoring (Week 8)

• limitwise_reference.csv: the Week 6 test set reweighted to the 3 percent book rate (every paying account four times, 578 defaulters once, 19,266 accounts), the file every month is read against. The Week 6 cut on this page is the score of the last account inside the top 10 percent of that file.
• limitwise_month_1.csv to _month_5.csv: five monthly files of 20,000 accounts, drawn at the book rate. One has no drift, one has data drift the model shrugs off, one has drift that moves the scores and the outcomes, one has a performance drop, one has a fairness gap. Tab 3 runs Evidently's data-drift report on any of them, or on a file of your own, and reads the caught rate and the groups the Week 6 memo watches.

The repositories

• Model repository (the model card, the signed file and its hash, the weights): https://huggingface.co/drdavidprivacy/limitwise-model
• Dataset repository (the datasheet and the remediation log): https://huggingface.co/datasets/drdavidprivacy/limitwise-data
• Portfolio page (the six deliverables as PDFs, the page each learner builds under their own name): https://huggingface.co/spaces/drdavidprivacy/corvana-portfolio
• Collection: https://huggingface.co/collections/drdavidprivacy/limitwise-governance-repository-6ae5ff9b374cab20b18250eb

What this page is not

A score says where a cardholder's record ranks on one list. It is not a decision about the cardholder. The decisions, and the conditions on them, are on the portfolio page.

Source

Built for the AI Governance Practicum (AI Governance Collective). The organization, the people, and the program are fictional. The data is a public dataset, remediated and documented in the dataset repository's datasheet.