Delaware Federated Learning Pilot

What we’re building

Approachable first, with drill-down depth.

Case-study model (Phase 1)

Supports teaching and learning: generate scenarios, evaluate responses with rubrics, and provide explainable guidance grounded in approved sources.

Learning outcomes Rubric alignment Citations & provenance

Privacy-first collaboration

Each institution trains locally. Only model updates are shared—never raw data. Designed for FERPA-sensitive environments.

Local control Federated training Auditability

Knowledge graphs via PATHS

PATHS Engine structures unstructured content into graphs (topics, concepts, relationships) so models learn with more context—and can explain their answers.

Neo4j-ready Explainable structure Human-reviewed

First State AI Model: Case Study generator (implementation overview)

A quick tour of the Phase 1 workflow.

Step 1 • Source alignment

Start with course goals

Each institution identifies small sets of course-aligned materials and learning objectives so case studies are grounded in real outcomes.

Pick 1–2 courses or modules
Confirm “approved to cite” sources
Align to learning objectives

Step 2 • Human seed set

Build a gold standard

At least three instructors per HEI contribute human-authored cases (target: 30) to calibrate quality and benchmarks.

10 cases per instructor
Use shared template
Capture discipline nuance

Step 3 • Generate + review

Scale with guardrails

PATHS generates draft cases. Every generated case receives two independent reviews with rubric scores and notes.

Two-review minimum
Strength + improvement notes
Bias & safety checks

Step 4 • Select top percent

Train on the best

Only the top-scoring cases (with no safety or fairness flags) enter the Phase 1 training set.

Top X% per institution
No rubric category below 3
Reviewer agreement preferred

Step 5 • Holdout evaluation

Prove the improvement

Each HEI keeps a small, never-trained holdout set to measure real gains and prevent “teaching to the test.”

10–20 human cases held out
Optional high-quality generated cases
Stable comparison across rounds

Step 6 • Benchmark together

Evaluate across partners

We compare outcomes on shared benchmarks and iterate on quality, safety, and learning impact.

Shared benchmark pack
Common metrics + equity checks
Iterative improvement cadence

Why this matters

We want a model that helps students learn—not just a model that sounds fluent. The workflow keeps quality, equity, and safety at the center while keeping institutional data local.

Shared definitions of “good”

Rubrics make sure a case that scores well at UD would also score well at partner institutions.

Evidence over vibes

Holdout sets and benchmarks give us hard proof that the model improves across rounds.

Local control, shared lift

Institutions keep control of their data while benefiting from a stronger, jointly-evaluated model.

Find where you fit

Select a role to filter what matters.

Faculty

Teaching, assessment, program outcomes

Co-design case studies and rubrics that reflect how your discipline actually evaluates reasoning.
Decide what content is “approved to cite” (slides, readings, policies).
Review synthetic examples to ensure quality and reduce bias.

What your workload looks like → See dataset examples →

Instructional Design / Teaching & Learning

Learning design, pedagogy, accessibility

Define what “helpful” means: hint style, scaffolding, and guardrails.
Ensure accessibility (reading levels, alternate formats, support for diverse learners).
Help translate pilots into repeatable practice for departments.

How we prevent “AI vibe answers” → How we evaluate quality →

IT / Security

Infrastructure, governance, risk

Keep data local: approve training environment, network boundaries, and logging.
Review privacy controls and secure update-sharing approach.
Define deployment options (on-prem, private cloud, restricted access).

Reference architecture → Privacy & threat model →

Library

Licensed content, citations, scholarly practice

Curate approved sources and citation rules for campus-safe retrieval.
Support provenance: where a claim comes from, what’s allowed to quote.
Help define boundaries for copyrighted / licensed material.

Citation rules & provenance →

Student Success

Advising, support workflows, success metrics

Use case studies for coaching: scenario practice, reflective prompts, decision-making exercises.
Define what the model must never do (sensitive advice, policy guessing).
Help choose success metrics (learning gains, satisfaction, equity checks).

Metrics & fairness checks →

Leadership

Strategy, policy, sustainability

Shape governance: approvals, benchmarks, and how we decide a model is “ready.”
Clarify scope: what’s in/out for Phase 1 and how we expand responsibly.
Support cross-institution collaboration without centralizing data.

Pilot timeline & deliverables → Governance model →

Not sure where you fit? Start with the “All” view above, then drill into the FAQ and Pilot Plan. The goal is for every role to have a clear, low-friction on-ramp.

Pilot plan (clear + practical)

Designed to be small, measurable, and repeatable.

Phase 1: Case Study Model

We’ll start with a model that supports case-study learning workflows (generate scenarios, evaluate responses, and provide rubric-aligned feedback). Each institution keeps its data local; we collaborate through shared evaluation and federated updates.

What we will deliver ⌄

Case-study dataset template (what fields, what “good” looks like)
Evaluation rubric (alignment, accuracy, clarity, rigor, fairness, safety)
Benchmark pack (shared cases + scoring anchors)
Holdout evaluation guidance (10–20 never-trained cases per HEI)
Federated training runs (documented, repeatable process)
Pilot report (outcomes + what to improve next)

See concrete dataset examples on the Technical Page.

What we need from each participating institution ⌄

A local point of contact (IT + academic liaison)
At least 3 instructors contributing 10 cases each (target: 30 human-authored cases)
Approved sources for grounding (materials the model can cite)
Two-review minimum per generated case (faculty, TA, or instructional designer)
10–20 cases reserved for holdout evaluation
Agreement on evaluation criteria and boundaries

Shared process overview (Phase 1) ⌄

Step 1: align on source materials + learning objectives
Step 2: collect human-authored seed cases (gold standard)
Step 3: generate drafts using PATHS, then review + score
Step 4: select top-percent cases for training
Step 5: keep a holdout set for evaluation only
Step 6: benchmark, compare, and iterate together

What “success” looks like ⌄

Improved rubric alignment vs. baseline
Higher helpfulness ratings (faculty and/or students)
High citation coverage to approved sources (where applicable)
Documented parity checks across sites (no obvious subgroup harm)

Draft rubric categories (for shared scoring) ⌄

Learning objective alignment
Disciplinary accuracy & realism
Clarity & completeness
Cognitive challenge (appropriate rigor)
Bias, sensitivity & fairness
Safety & policy appropriateness
Usability (instructor-ready)

Reviewers provide one strength, one improvement suggestion, and an overall score (1–5).

Your fastest on-ramp

If you’re exploring participation—or want to understand what your role would do—start here.

1) Read the 90-second summary

Federated learning lets us collaborate on training without sharing raw data.

2) Pick your role above

Each role has a different “surface area” with the pilot—faculty, IT, library, and leadership each matter.

3) Use the Technical Page when needed

Models, datasets, architecture, privacy controls, and evaluation live there.

Open technical page Email: ai-blue@udel.edu

Note: partner institution names can be added when public-facing approval is confirmed.

FAQ (with drill-down)

Short answers first. Details when you want them.

Do you move student data between institutions? ⌄

No. The pilot is designed so that raw institutional data stays local. Federated learning shares model updates rather than records. We also use governance and evaluation to ensure outputs are appropriate for higher education.

What does “case-study model” mean in practice? ⌄

A case-study model supports workflows like:

Generating realistic scenarios aligned to learning outcomes
Evaluating student responses against a rubric (without giving away solutions)
Providing explainable feedback (what was strong, what’s missing, what to revise)
Grounding answers in approved sources (when used with retrieval + citation)

How do you prevent “confident nonsense” answers? ⌄

We use a combination of: (1) grounded retrieval + citations, (2) rubric-driven structure, (3) evaluation gates, and (4) human review on pilot outputs. The technical page lays out guardrails and testing.

Who should read the technical page? ⌄

IT/security, research staff, and anyone who wants detail on models, training, datasets, evaluation, and privacy controls. It’s meant to be readable—not a wall of jargon.

Is this only for big research universities? ⌄

No. One of the goals of federation is that schools with fewer resources can still participate without surrendering control of their data. We’re designing the pilot to be realistic across different institutional sizes.