Start from one concrete question
What changes if we remove one example, swap in a different subset, or let a group of contributors walk away?
Training-data field guide
See how changing data changes the model.
Most machine-learning discussions treat training data like a hidden blob in the background. This site does the opposite. It turns training data into something you can inspect, compare, remove, corrupt, and reason about directly.
The core move is simple: ask a concrete what-if question about the data, then trace how model behavior changes. That one move links influence functions, scaling laws, data valuation, poisoning, and collective action without flattening them into buzzwords.
- Interactive toy grid
- Guided examples
- Curated paper collections
- Glossary for fast orientation
Use the same visual language throughout
The memo, glossary, paper lists, and explorer all point back to the same grid metaphor instead of inventing fresh jargon on each page.
Move from intuition to literature
Once the idea clicks, you can jump directly into curated references for the subfields that matter most to you.
The basic move
Compare nearby worlds, not vague abstractions
The easiest counterfactual is leave-one-out: compare a training set that includes one point with the same training set after that point is removed.
That sounds almost too simple, but it gets you surprisingly far. From there you can ask the same question about groups, subsets of a fixed size, corrupted examples, or withheld data.
- Rows tell you what changed in training. Move between rows to simulate different data choices.
- Columns tell you where the effect lands. A point may matter a lot for one evaluation slice and barely at all for another.
- Differences are the real story. A single score is rarely interesting on its own; the comparison is what matters.
Leave-one-out example
ABCD versus ACD
Here the gold row trains with B and the teal row leaves B out. The biggest drop lands on evaluation point B, which is the intuition people often want in plain language.
A
B
C
D
ABCD
0.92
0.88
0.85
0.82
ACD
0.78
0.55
0.82
0.80
This is the same intuition behind influence estimates and many forms of data attribution: compare a baseline training world with a nearby altered one.
What this framing connects
Different literatures, same underlying move
The point of the project is not to pretend every subfield is identical. It is to show that many of them start from a shared operation: change the data, then compare the outcome.
Influence and attribution
Compare nearby rows to see which examples, or which groups of examples, are doing the work.
Scaling and selection
Scan the grid by subset size to ask when more data helps and which subsets punch above their weight.
Poisoning and robustness
Treat corrupted or withheld data as another counterfactual and inspect the damage directly.
Collective action
Ask what happens when many contributors walk away together, instead of pretending each point acts alone.
Why people care
Useful for debugging, bargaining, and model hygiene
Debug a model failure
When a prediction looks wrong, the framework gives you a concrete way to ask which training examples changed the outcome.
Value a dataset
Instead of talking about data value in the abstract, you can compare outcomes with and without a slice of the data.
Reason about leverage
People who contribute data are not just passive inputs. This framing makes their bargaining power legible.
Where to go next