Teacher field guide

Start with the research demonstration. Use the notebook to test a specific claim with a smaller model, then discuss what the evidence can and cannot establish.

Choose a classroom route

Activity	Hosted browser notebook	Downloaded native notebook
Years 10 / 11 · Calculation	Train a small neural network; compare training and reserved pairs.	Run SmolLM2; probe real hidden states, test time transfer and a controlled intervention.
Year 6 · Factuality	Run a trained tiny transformer, learn a fallible checker, train an answer policy and measure mistakes plus coverage.	The same experiment; editable Python.
Years 5 / 11 · Stories	Paste or upload a short story; compare real MiniLM encoder vectors for two endings.	Collect SmolLM2 hidden states; optionally score emotion labels. Year 11 can import and select Gutenberg passages.

The original Goodfire embeds remain available independently of our model downloads. Use their full-window links on smaller screens. Their controls and accessibility are maintained by Goodfire; Spark’s own outputs also have text tables.

Run a downloaded notebook

Use a teacher-managed computer with Python 3.12 or newer. Download the .py notebook from its lesson. Create an environment once, then open the file in Marimo.

python -m venv .spark-venv
# macOS / Linux:
source .spark-venv/bin/activate
# Windows PowerShell instead:
# .spark-venv\Scripts\Activate.ps1
python -m pip install "marimo==0.24.0" "numpy>=2,<3" "transformers==4.57.6" "torch>=2.6,<3"
marimo edit spark-y11-stories.py

Choose native SmolLM in a story notebook, or submit the pretrained-model form in a calculation notebook. The model is downloaded only when you run that experiment. SmolLM2-135M’s weights are about 270 MB; libraries add more. Allow several minutes and roughly 2 GB of working memory. The default runs on CPU.

These small experiments do not need an H100. If your school already has an authorised CUDA runtime, install its compatible PyTorch build and set SPARK_DEVICE=cuda before opening Marimo. This uses that existing GPU; the site does not provision or charge for a GPU. CPU execution has been checked; CUDA is an optional extension. See the existing hardware guide for the wider Brightlab setup.

To study the exact arithmetic research model and methods, use Goodfire’s accompanying research repository. Spark’s small-model probes are an extension, not a reproduction of the Llama results.

Stories your class can read

Year 5: read an original Goodfire example together, then use a short class-written story. Replace one ending. Ask students to point to the sentence that changed and explain a different possible interpretation.

Year 11: import Hamlet or Grimms’ Fairy Tales in native Marimo. Choose paragraphs, preview them, and edit to 3–40 sentences before analysing. Record the edition and boundaries. Dialogue and stage directions make automatic sentence splitting imperfect; inspect the table.

A full book is not one model context. MiniLM keeps the first 256 tokens; the native LLM keeps the last 512. Truncation is reported for both versions. Select a shorter passage when the comparison loses relevant text. Teachers should check the chosen edition’s licence and passage suitability.

Assess the investigation

Look for four things: a prediction before the run; a comparison with one deliberate change; measurements used accurately; and a conclusion that acknowledges a limitation.

Years 5–6 can explain aloud, annotate a printed table or draw their reasoning. Years 10–11 should submit the downloaded record with model revision, settings, reserved examples and a proposed control. These are research enrichment activities; they are not an independently certified curriculum alignment or graded assessment package.

Research distinctions to preserve

A Fourier circle drawn from a formula is not evidence that a model learned that circle.
A probe finding information does not prove the model uses it. An intervention can also introduce disruption.
Correcting one answer does not update model weights. Training from feedback can change future behaviour, but a fallible reward can teach mistakes.
Fewer wrong answers must be considered alongside how often the model answers.
Emotion readouts, hidden states and a two-dimensional projection are three different measurements.
Goodfire’s RLFR headline improvement combines several techniques; it is not the effect of feedback training alone.

School devices, privacy and saved work

Hosted Python and models load from marimo’s runtime services, jsDelivr, Python package hosting and Hugging Face. Goodfire demonstrations load from static.goodfire.ai. If downloads are blocked, use the native notebook on a teacher-managed device. No account, model API key or paid model service is needed for the small-model activities.

Our analyser computes student text locally. Remote hosts receive ordinary download requests, not the story text. Use fictional or public-domain material; avoid student personal information. Reloading clears the notebook session. Save the evidence and reflection before closing it.

The hosted site is publicly accessible at brightlab.education. A downloaded notebook or evidence file can be shared separately by its holder.