Control tokens for slides

In part 3 I bolted a control knob onto the storyteller: dictate a shape token before each sentence and the story's structure follows. The knob worked but was blunt — story shape is an emergent property, so you could suppress dialogue reliably but not force it past what the plot afforded.

That raised a question. What if the thing you're controlling isn't emergent, but something the model writes down explicitly? To test it I needed a different task, so I built one: a local slide generator, and then gave it a control token for slide layout. The result is the cleanest steering in the whole series.

A local slide generator, first

Presentation tools turn a document into slides by calling a cloud model on each chunk. I wanted that offline. So I distilled it: I ran an existing cloud slide-maker (gpt-5.4-mini) over 2,500 text chunks and kept the (chunk → slide JSON) pairs it produced — each slide a small object with a layout, a title, and bullets. Then, exactly as in part 1, I trimmed Gemma's vocabulary to English (28k tokens → a 118M model) and fine-tuned it (LoRA) on those pairs.

It runs fully offline and produces valid slides. As always at this size, the format is nailed and the content is imperfect — occasionally a garbled title or an inverted fact. Form is cheap; content is hard. (A recurring tune on this blog.)

The control token: layout

Every slide the teacher produced already carried a layout field, one of four: bullets, statement, quote, stat. That's a free control label for every training example — I didn't have to annotate anything. So I added four tokens to the vocabulary, one per layout, and retrained with the desired layout token prepended to each passage:

Make one slide as JSON.
LAYOUT: <quote>
PASSAGE: ...
SLIDE JSON:

At training the token is the layout the teacher actually chose. At inference I dictate it — "make this a quote slide" — and read back what the model produced.

It obeys, almost perfectly

I forced each of the four layouts on 80 held-out passages and checked the generated slide's layout:

forced ↓ / got →	bullets	statement	quote	stat
bullets	79	·	·	·
statement	2	78	·	·
quote	·	·	79	·
stat	·	·	·	77

Mean adherence 0.98, against a chance rate of 0.25. Tell it "quote" and you get a quote slide 99% of the time.

This is far cleaner than story shape, and the reason is the interesting part: layout is something the model writes down explicitly — it literally emits "layout": "quote" — whereas story shape had to be inferred from the sentences it generated. When the controlled attribute is an explicit output field, the control token lands almost perfectly. When it's emergent, control is real but partial. Same mechanism, very different ceilings.

What the knob can't do: make a statistic exist

Near-perfect adherence doesn't mean near-perfect slides. Force a layout the passage can't support and the format holds while the content strains. Here's all four layouts forced on a passage about a lighthouse keeper — no numbers anywhere:

• bullets → "helped hundreds of ships" ✓
• statement → "saved hundreds more ships" ✓
• quote → lifts the passage's own quote-like line ✓
• stat → format-correct, but its value field comes out as "years" with the label "saved ships over decades" — there is no statistic to render, so the model fabricates one to fill the slot.

That's the ceiling, and it's the same one stories hit with dialogue: the control token dictates the form, and the failure mode is always content, never format. The model does exactly what you asked structurally, and pays for it in words.

The through-line, one more time

Two modalities now tell the identical story. Trim the vocabulary and the model shrinks for free. Rent fluency from pretraining and a few thousand examples teach it a task. Add a control token — a shape for stories, a layout for slides — and you get a cheap, reliable, measurable handle on structure: 0.98 for the explicit attribute, partial for the emergent one. And in both, the knob turns form, not quality. The good sentence and the real statistic still have to come from somewhere else.

This pair of results — narrative shape and slide layout — is written up as a short standalone paper in the project repo.