Skip to main content

Stage 2: Train Your First Model

Course progressStage 2 of 10
~60 min
Your two tools

Keep both tabs open all week. Open in a new tab — don’t use the buttons in this page to leave the course.

Train your AITeachable MachineOpen in a new tab ↗Build the gameRAISE PlaygroundOpen in a new tab ↗
Build

lots of example photos in every class, then a trained model

Learn

what happens inside the AI when we press the Train button

Ship

a model that can guess your hand sign live, with a URL you can save

Teacher demo

Before campers start, show the room:

  1. Open your own Teachable Machine project. Fill Rock with about 15 photos by holding the Hold to Record button for 4 seconds.
  2. Show what happens when you move your hand a little while recording — many slightly different photos appear.
  3. Click Train Model. Wait. Then put your hand in front of the camera and let the room see the live guess.
  4. Click Export Model → Upload my model → Copy the shareable link. Show campers exactly where the URL lives.

The big idea​

Today the AI starts to learn. We are going to fill up all four classes with photos, click the Train button, and watch what happens.

[Rock photos]    [Paper photos]    [Scissors photos]    [Nothing photos]
       │               │                   │                    │
       └───────────────┴───────────────────┴────────────────────┘
                                 │
                                 ▼
                         🟢 Train Model 🟢
                                 │
                                 ▼
                A model that can guess what you show it

A model is the AI's brain after it has practiced. Before training, the AI is empty. After training, the model has seen all the patterns in your photos and can guess what new photos belong to.

The most important rule of training: more photos = better guesses. We aim for about 15–20 photos per class, taken with small differences (slightly different angles, slightly different distance from the camera). Variety is what makes a model strong.

New words
train
let the AI practice with the examples
model
the AI's brain after it has practiced
prediction
the AI's guess about what it sees right now
URL
the web address where your model lives online
Before you start

Your Stage 1 Teachable Machine project should still be open with four classes named Rock, Paper, Scissors, and Nothing.

Build it​

Step 1 — Fill every class with photos​

Aim for about 15 photos per class. Don't worry about counting exactly — the bar inside each class fills up so you can eyeball it.

For each class in order (Rock → Paper → Scissors → Nothing):

  1. Click the Webcam button in that class box.
  2. Make the hand sign for the class. (For Nothing, just hold an empty hand or no hand at all.)
  3. Press and hold the Hold to Record button for about 4 seconds.
  4. While you hold the button, slightly move your hand: tilt it, lean a little to the side, move closer and farther from the camera. Variety matters.
  5. Let go. Look at the photos. If they all look identical, take 5 more with more movement.

Repeat for all four classes.

Step 2 — Click Train Model​

On the right side of the screen, find the green Train Model button. Click it.

A spinning wheel appears. Don't close the tab. Don't click anywhere else. Training takes about 30 seconds to a minute. The wheel will show Listening for examples... and then Training...

When it's done, the Preview panel on the right side will start showing bars for each class.

Step 3 — Test your model live​

Make a rock hand sign at the camera. Look at the Preview panel.

You should see a tall green bar next to Rock — that's the AI's guess, with a number showing how sure it is. Smaller bars next to other classes show how much the AI considered those options.

Test all four signs:

  • Rock — closed fist
  • Paper — flat open hand
  • Scissors — two fingers in a V
  • Nothing — hand out of frame

If three out of four work, you have a working model.

Step 4 — Save your model URL​

This URL is the bridge to Stage 4, when we move the AI into the game. Don't skip this step.

  1. Click the Export Model button (top-right of the Preview panel).
  2. A dialog opens. Click Upload (shareable link).
  3. Click Upload my model. Wait for the upload — it can take up to a minute.
  4. A long link appears. It looks like: 
    https://teachablemachine.withgoogle.com/models/AbC123xyz/
  5. Copy that link. Paste it where your coach told you to save it (Google Doc, sticky note, notebook).

The model URL is now saved. You can come back to it any time.

Understand it​

Training is the AI finding patterns across all your photos. The AI doesn't see "rock" or "scissors" — it sees pixels. But when 15 photos labeled "Rock" all have a roundish shape in the middle, the AI starts to learn that roundish shape in the middle = probably Rock.

This is why variety matters. If every Rock photo is in the exact same spot, the AI learns "Rock = something at that exact spot" — and breaks the moment you move your hand. With varied photos, the AI learns the actual shape, not the spot.

The Preview bars show confidence. A 99% bar means "I'm almost sure." A 60% bar means "I think so, but I'm not certain." Two close bars (like 51% / 49%) mean "I really can't tell." Watching the bars is how you'll spot the AI's weak spots in Stage 3.

The URL matters because Teachable Machine and RAISE Playground are two separate websites. They can't see each other directly. The URL is how RAISE Playground (in Stage 4) will know which AI brain to load.

Try this​

Learning beat

Try this

Three short experiments. Predict before you run, then test your guess.

Predict first

Imagine you only took 3 photos per class instead of 15. Predict: will the model still train? Will the bars still appear? Will it guess correctly? Why or why not?

Compare

With your model running, hold your hand right up close to the camera, then far away. Do the bars stay the same? What does that tell you about how the AI thinks of "Rock"?

Connect

Stage 3 finds spots where your model gets confused. Watch the Preview bars right now. Is there one pair of classes the AI mixes up more than the others? (Hint: Rock and Nothing are common.)

Test your stage​

  • Every class has roughly 15+ photos.
  • You clicked Train Model and the wheel finished without errors.
  • At least 3 of the 4 hand signs make the right green bar light up.
  • Your model URL is saved somewhere outside Teachable Machine.
  • Design check. Hold each hand sign for 5 seconds. Does the bar stay steady, or does it flicker between two classes? Flickering means the model is unsure — Stage 3 will fix that.

If it breaks​

  • The Train button is grayed out. One of your classes has zero photos. Look at all four boxes — every one needs at least a few photos before training can start.
  • Training takes forever. This usually means a lot of photos in one class. If a class has 200 photos and another has 15, the AI is overworking. Aim for roughly equal counts.
  • The model says everything is "Nothing". Your Nothing class probably has way more photos than the others. Delete some Nothing photos by clicking the trash on individual ones, or take more photos for Rock / Paper / Scissors.
  • The URL upload fails. Refresh the page, retrain (your photos are still there), and try Export Model again. The upload step needs a working internet connection.
  • I forgot to save my URL and the tab closed. Your photos are still saved if you used the same browser. Reopen Teachable Machine, your project should still be there. Train again and re-export.
Coach notes

The most common failure: kids hold the Hold to Record button perfectly still for 4 seconds. They get 30 nearly-identical photos. The model then overfits to that exact pose and breaks when their hand moves. Walk the room during photo capture and gently remind: "wiggle a little while you hold the button."

Equally common: skipping the URL save. Without the URL, Stage 4 is dead. Build a moment into the timing — "raise your hand when your URL is saved, I'll come check it" — before letting anyone move on.

For a fast group, the medium stretch (lighting variations) is the highest-value addition. The hard stretch (5th class) is fun but can spiral into "add ten classes" — cap it at one extra.