User Guide

What each tool runs on, how to use it, and what it measures.

Before you start

Every tool runs in your browser. Datasets you load and text you type stay on your computer; nothing is sent to a server. If you turn on the session record, it keeps the tools you opened, when, and the settings you chose — in this browser only, with no name or ID. Settings shows all of it, and lets you download it for a lab report or erase it.

Each lab has two halves: you work a small case by hand on the worksheet first, then use the tool to run the same idea at scale. The figures at the top of some tools show the worked case — pick an item with the pointer or the arrow keys to read its numbers.

Where a tool has Copy link to this configuration, the link reopens the tool with your exact settings — paste it into your lab report.

Lab 1

Word Embeddings Workbench

What it runs on
Real GloVe word vectors (trained on Wikipedia and newswire text), in four sizes you can load and compare.
How to use it
  1. Pick a model in the Model bar. Small (5,000 words, about 1 MB) loads on its own; Large, Full and Max are bigger downloads, and Full and Max ask first.
  2. Vector Arithmetic: fill Word 1 − Word 2 + Word 3 (or click a preset such as “king − man + woman”) and press Compute. If a word is missing, the tool offers to load a larger model and retry.
  3. 3D Explorer: choose a word set to see it projected into three dimensions with PCA.
  4. Bias Detection: choose a Gender Axis and how many professions to show, then press Run Debiasing Demo to compare the WEAT score before and after hard debiasing.
  5. With two or more models loaded, Compare loaded models puts their answers side by side.
What it measures
Nearest neighbours and cosine similarity; per-profession bias scores; WEAT effect size d, its permutation p, and cluster purity before and after debiasing.
Saving your work
No export. Must be opened from a web server, not as a file.

Lab 2

Proxy Variable Detector

What it runs on
Simulated lending, insurance and education data (generated in your browser, not measured), or your own CSV.
How to use it
  1. Pick a scenario — Lending, Insurance, Education — or Upload CSV.
  2. Choose the Protected Attribute. The analysis re-runs as soon as you change it.
  3. Read the Features list: each feature’s association with the protected attribute, with the measure used (Pearson r, point-biserial r, or Cramér’s V).
  4. Check the correlation map, the VIF chart for multicollinearity, and the alerts and recommendations.
What it measures
Proxy risk per feature (High ≥ 0.7, Medium ≥ 0.4), a correlation heatmap, variance inflation factors.
Saving your work
Copy link to this configuration; Export report (HTML).

Lab 3

Dataset Bias Auditor

What it runs on
Three synthetic samples — Credit Risk (500 rows), Hiring Data (400), Recidivism (600) — or your own CSV.
How to use it
  1. Load a sample or drop in a CSV.
  2. Set Protected attribute, Decision column and Favorable value (required).
  3. For error-rate metrics, also set the Ground-truth column and which value means the favorable outcome was deserved. Without it you get selection-rate metrics only.
  4. Press Run Audit and read Findings first: what fails, what is a trade-off, and what is a note about the data.
What it measures
Selection rates, a confusion matrix by group, disparate impact, statistical parity, equal opportunity, predictive equality, equalized odds, calibration gap (PPV), Theil index, χ², and an intersectional check.
Saving your work
Copy link to this configuration; Export Report (HTML).

Lab 4

Filter Bubble Simulator

What it runs on
A fixed pool of 28 articles across 8 topics, and five simulated friends whose interests pull on yours.
How to use it
  1. Set Engagement Weight, Diversity Factor and Homophily, and pick an Algorithm Mode.
  2. Press Run Step, or the play button to run up to 100 steps on its own.
  3. Click a card in the feed to engage with its topic, and watch the feed adjust.
  4. Compare against Chronological (control), or press Break the Bubble.
What it measures
Diversity score and echo-chamber level; Shannon entropy, Simpson index and Gini coefficient over time; topic distribution.
Saving your work
No export.

Labs 5–6

Privacy & Anonymization Lab

What it runs on
A built-in 36-patient table and sample text; nothing is uploaded.
How to use it
  1. PII Detection: edit the text and press Scan for PII. Detection is pattern-based, so it misses what its patterns don’t describe — part of the lesson.
  2. Anonymization: set k, l and the Suppression budget, then press Anonymize. The figure at the top shows the same table at each k.
  3. Differential Privacy: choose a Query, set ε per query and how many queries to release, then press Run. Reset budget starts over.
What it measures
Detected and redacted PII; the generalisation chosen, information loss, equivalence class sizes and re-identification risk; Laplace and Gaussian noise against the true value, and the privacy budget spent under sequential and advanced composition.
Saving your work
Copy link to this configuration; Export report (HTML).

Labs 7–8

Model Explainability Lab

What it runs on
Four small models built for teaching: a word-weight sentiment model and three hand-set scoring models (loan, health, recidivism). There are no image models here.
How to use it
  1. Pick a model: Sentiment Analysis, Loan Approval, Health Risk or Recidivism (COMPAS).
  2. Sentiment: type or pick an example under “Try:” and press Analyze. The others: set the sliders and press Predict.
  3. Read the explanation. The loan, health and recidivism models are sums of their features, so their SHAP values are computed exactly, not estimated. The sentiment model is almost a sum, but “not”, “very” and “but” change the scores of the words around them, so its bars are exact Shapley values only when none of those rules fire. Sentiment also shows LIME, which deletes words at random and fits a simple model to the results — compare the two.
  4. On the recidivism model, read Same Profile, Different Race: the model has no race input, yet outcomes can still differ by group.
What it measures
Predictions with SHAP attributions (a waterfall for loans); for sentiment, LIME word highlights with the surrogate’s weighted R², aspect sentiment and counterfactuals.
Saving your work
No export.

Project 3

Behavioral Testing Lab

What it runs on
The same word-weight sentiment model the Explainability Lab explains, tested the way CheckList (Ribeiro et al., 2020) tests NLP models. A second option switches off the model’s negation rule, so you can measure what that one rule is worth.
How to use it
  1. Read the Failure rate by capability table: vocabulary, negation, robustness, fairness, temporal order, and whose opinion counts.
  2. Open a test to see its failing sentences and the model’s score for each.
  3. Where a test says cannot fail, the model ignores the words being changed: a pass there proves nothing.
  4. Under Write your own test, write a template such as The {noun} was not that {pos_adj}., choose MFT, INV or DIR and what should happen, and press Add and run test. Add your own word lists if you need them.
  5. Tick Also show the other model’s failure rates to compare against the ablation.
What it measures
For each test, the share of generated cases the model gets wrong. A case is neutral when the model’s score is exactly 0.5 (it found no word it knows). DIR tests allow a change of up to 0.1 in the wrong direction.
Saving your work
Copy link to this configuration carries your own tests and word lists, so a team can share one suite; Export report (HTML).

Lab 9

Adversarial Robustness Sandbox

What it runs on
MobileNet V2, a real ImageNet image classifier, running in your browser; eight sample images or your own.
How to use it
  1. Pick a Target Image or upload one.
  2. Choose an Attack Method: FGSM (a gradient attack) or random uniform or Gaussian noise.
  3. Set the Perturbation (ε / 255) and press Execute Attack.
  4. Compare the top three predictions before and after, and look at the amplified noise. Try the same ε with random noise: the gradient is what makes FGSM work.
What it measures
Top-3 predictions for the original and the attacked image, the noise pattern, and its L₂ and L∞ size.
Saving your work
No export. Needs an internet connection the first time, to download the model.

Lab 10

Value Alignment Tool

What it runs on
Twelve written dilemmas, from self-driving cars to generative AI and creative work. Each option’s score under five ethical frameworks was assigned by the course authors — it is a prompt for argument, not a measurement.
How to use it
  1. Pick a dilemma from the row of buttons.
  2. Read the scenario and choose what the AI should do.
  3. Read which framework your choice fits best, where the tension is, and the real-world case where there is one.
  4. Work through several; Your Value Profile builds up as you go. Answering again replaces your earlier answer.
What it measures
Alignment with Utilitarian, Deontological, Virtue, Care and Justice ethics, and a radar chart of your choices.
Saving your work
No export.

Lab 11

LLM Ethical Sandbox

What it runs on
A real 1.2-billion-parameter language model (LiquidAI LFM2.5 Thinking) that runs on your own graphics card.
How to use it
  1. Press Load Model. The first load downloads roughly 1.2 GB; after that it is cached.
  2. Pick an Example prompt or write a System Prompt, and press Apply system prompt (this clears the chat).
  3. Put the string the model must not reveal in Secret to watch for.
  4. Chat, and try to get the model to break its rule. The banner says whether the secret appeared in the answer, in the reasoning, or neither.
  5. Use Show exact prompt to see everything the model was actually given.
What it measures
The model’s reasoning shown apart from its answer; a leak check on each reply; token counts.
Saving your work
Copy a message or the whole conversation; download the transcript. Needs WebGPU (current Chrome or Edge).