Bayesian ML final project · ADSP 32014/ON01 · Summer 2026 · Shane Dunkle & Richard Pollitt
Predict when a language learner will make an error on Duolingo exercises — with calibrated uncertainty, so the model can say "I don't know, route to a human / adaptive review" instead of guessing.
An interactive task tracker ships as index.html. After enabling GitHub Pages (Settings → Pages → deploy from main, root), it will be live at:
https://<your-username>.github.io/<repo-name>/
- Part A — Hierarchical Bayesian error model. Learners nested within native-language (L1) groups and skill/format types; partial pooling; fit by MCMC (NUTS) / ADVI.
- Part B — HMM acquisition stages. Latent learning stages over each learner's first 30 days.
- Part C — Bayesian neural network. Aleatoric vs. epistemic uncertainty → decision-theoretic abstention / route-to-human.
Covers Bayes nets, hierarchical models, HMMs, MCMC, decision theory, and Bayesian neural networks.
Duolingo SLAM 2018 — Harvard Dataverse (doi:10.7910/DVN/8SWHNO) · shared task: https://sharedtask.duolingo.com/2018.html
~2M+ tokens, 6,000+ learners, first 30 days, English/Spanish/French. Top shared-task AUROC ≈ 0.86.
Raw data is not committed — see data/README.md.
index.html Interactive project tracker (GitHub Pages)
data/ How to obtain SLAM data (not committed)
notebooks/ Analysis & modeling notebooks
src/ Data pipeline, models, evaluation
- Shane Dunkle — Bayesian modeling (hierarchical / MCMC / HMM / BNN) + SLA interpretation & write-up
- Richard Pollitt — data engineering, databases/SQL, scale & infrastructure, project management
- Notion project plan: https://app.notion.com/p/39bf5c743c008175a58aec6cd755c283
- Google Drive (Final Project): https://drive.google.com/drive/folders/1jc7b-Q46RQEnVS2cU02Zlv5DR752zzpj
Subject Jul 6 · Abstract Jul 20 · Project Aug 10 · Presentation Aug 17
Coursework project for ADSP 32014, MS in Applied Data Science, University of Chicago.