presentations/README.md

# Presentations

Talks about FinePhrase, one folder per talk (`YYYY-MM[-DD]-<venue>`). The COLM 2026 poster lives in `../poster/`.

| Folder                                 | Talk                                       | Length  | Slides |
| -------------------------------------- | ------------------------------------------ | ------- | ------ |
| `2026-02-26-se2026-industry-day`       | SE 2026 Industry Day, Bern (reveal.js)     | 30 min  | 21     |
| `2026-06-18-deep-learning-with-yacine` | Deep Learning with Yacine (full deep dive) | ~50 min | 80     |
| `2026-10-colm-spotlight`               | COLM 2026 spotlight oral, San Francisco    | 12 min  | 19     |

## SE 2026 Industry Day

`index.html` is a [reveal.js](https://revealjs.com/) deck whose D3 charts (`charts/`) load `data/` at runtime, so serve the folder over HTTP to present it (`python3 -m http.server`). `build-standalone.mjs` inlines the CSS, images, charts and data into `standalone.html`, a single file you can open straight from disk (reveal.js and D3 still come from a CDN):

```bash
cd presentations/2026-02-26-se2026-industry-day
node build-standalone.mjs
```

## COLM 2026 spotlight

`build.mjs` generates `finephrase-colm-2026-spotlight.pptx` with [pptxgenjs](https://gitbrent.github.io/PptxGenJS/). All charts are native PowerPoint charts with scores from the paper (`paper/colm2026_conference.tex`, `paper/data/*.dat`), except the trained-model-size slide, which uses the blog's student sweep (`app/src/content/assets/data/student-scaling-summary.csv`), and the headline-result slide, which shows the training curves of the blog's opening figure (`app/src/content/assets/data/benchmark-results.csv`). Every slide has speaker notes written to be read aloud, each starting with a target duration like `[45s]` (about 8.5 minutes in total). The last slide's notes collect material for questions that is not meant to be read aloud: seed variance (from the blog's 72-run seed grid), the baseline scores when mixed 50/50 with web data, cost comparisons, a training-compute estimate per model size, and a one-line description of each benchmark. `assets/` holds the paper overview figure, the bookshelf and dataset-viewer images from the Yacine deck, and the QR codes from the poster. The closing slide invites people to the poster session and shows `../poster/finephrase-poster.png`, so rebuild the deck after changing the poster.

```bash
cd presentations/2026-10-colm-spotlight
npm install
node build.mjs
```