← Back to course
AI and You 9-12 / Week 03 / Friday
5/6
Week 03 · How Models Learn

Friday

Impact Friday: What did SH-1 learn from?
// The crew trains Tiny Text by hand to see how training data becomes predictions
⏱ about 20 min

Friday: Impact Friday: What Did SH-1 Learn From?

Comet spreads the packet open to P3 again. "We know exactly what Tiny Text learned from," she says. "Six cards. But SH-1 just says public text and study material."

Wren writes two headings on his clipboard: Known and Unknown. "What does the evidence say? The Unknown column is getting long."

Teacher Mara looks over his shoulder. "Was any of that study material in Spanish? Was any of it from classes like ours?"

Nova projects the crew's model card with a new, empty section. "Would you like a hint?" she asks. "Write down only what the packet actually says."

Comet nods slowly and hands you the pencil. "Reviewer, fill in what we know and what we don't."

What developers share about training data

Training generative AI takes large amounts of data. In some cases, that data may include personal data.

Most model developers do not say exactly which data sources they trained on.

So "public text and study material" may be all the builder knows about the base model, or all it chose to share.

Who training data serves

Biased training data can leave some people less well served than others.

Tiny Text shows a small version. It has three library cards and only one gym card.

If students asked Tiny Text about the gym, it would have far less to go on. A made-up example, but the idea is real.

WHAT CAN WE KNOW?
  • Read the question.
  • Tap your answer.
Packet section P3 lists no sources. What does that tell the crew?
Could SH-1's training data include personal data?
Tiny Text has three library cards and one gym card. Who is less well served?
AI actor spotlight
AI design: gathering, cleaning and documenting data.
AI design work writes down a system's concept, goals, assumptions, setting and requirements. It gathers and cleans data and documents the data.
People who do it include data scientists, domain experts, members of affected communities, user-experience designers, data engineers and privacy experts.
In the SH-1 review, the crew is asking the builder for exactly what this work should produce: a record of the data.

Model card, part 3: training data

Known (from the packet)Unknown (questions for the builder)
"Public text and study material" (P3)Which public text, and which study material?
Built on a large pre-trained language model from another company (P2)What the base model was trained on
Tuned on study material (P2)Whether any of it came from classes like Harbor Point's, or was in Spanish
No list of sources was givenWhether it includes personal data, and how old it is

Week review

  1. Machine learning is statistical inference that finds patterns in data, and it needs huge amounts of training data.
  2. Language models predict the next word. A model can be a table of patterns.
  3. Training data can go stale, and changes during training change how a system performs.
  4. Training data can include personal data, and most developers do not list their sources.
  5. Biased training data can leave some people less well served.
Week reviewTrue or false?
Most model developers list exactly which data sources they trained on.?
Training data for generative AI can include personal data.?
Biased training data can leave some people less well served.?
Tiny Text's tallies come only from its training cards.?
Documenting the data is part of AI design work.?
WHY THIS EXERCISEWhat a model learns from shapes who it serves, so a reviewer always asks about the data.
OUR WEEK, IN ORDER
  • ?We counted next-word pairs in three Tiny Text cards.
  • ?We found a pattern in Teacher Dev's labelled questions.
  • ?We retrained Tiny Text when the library hours changed.
  • ?We trained Tiny Text on all six cards.
  • ?We listed what we know and do not know about SH-1's training data.
WHY THIS EXERCISEEach step showed one more way training data shapes a model.

Wonderful work, reviewer. Tomorrow's Explorer Quest trains a Tiny Text at home.

← Thursday