|
1 | | -# 🎥 Screencast script & storyboard: "Decanting the Web" |
| 1 | +# 🎥 Screencast: "Decanting the Web" |
2 | 2 |
|
3 | | -**Target length:** ≤ 2:00 (hard limit). **Aim for 1:50** to leave a safety margin. |
4 | | -**Goal (per rubric):** show what the viz *does* in a fun, engaging, impactful way, and talk about **contributions and insights**, *not* technical details. |
| 3 | +**▶ Watch the 2-minute video:** |
| 4 | +https://drive.google.com/file/d/1bBN4NYmE1X15sft0V5nxJhfG020pXn0W/view?usp=sharing |
5 | 5 |
|
6 | | -> Tips: record at 1920×1080, hide the browser bookmarks bar, use a clean profile, and |
7 | | -> pre-scroll once to warm up fonts/animations. Scroll **slowly and smoothly** (a trackpad |
8 | | -> or a scroll-automation tool helps). Record narration separately and lay it over the |
9 | | -> screen capture for clean audio. |
| 6 | +A two-minute narrated walkthrough of the interactive visualization for a general, |
| 7 | +ML-curious audience. It follows the data story end to end: the curation funnel, document |
| 8 | +length, the domain galaxy, domain concentration, and the quality landscape, with the |
| 9 | +interactions demonstrated live. The beat sheet below documents the structure of the video. |
10 | 10 |
|
11 | 11 | --- |
12 | 12 |
|
13 | | -## ⏱️ Beat sheet |
| 13 | +## Beat sheet (structure of the video) |
14 | 14 |
|
15 | | -| Time | On screen | Narration (read aloud) | |
16 | | -|------|-----------|------------------------| |
17 | | -| **0:00-0:10** | Hero section. Let the title animation breathe; cursor still. | "Every AI model is what it eats. Before it can write, it has to read, and most of what it reads comes from one dataset: **FineWeb**. This is the story of how the open web becomes AI training data." | |
| 15 | +| Time | On screen | Narration | |
| 16 | +|------|-----------|-----------| |
| 17 | +| **0:00-0:10** | Hero section. The title animation settles. | "Every AI model is what it eats. Before it can write, it has to read, and most of what it reads comes from one dataset: **FineWeb**. This is the story of how the open web becomes AI training data." | |
18 | 18 | | **0:10-0:22** | Click **"Follow the journey"**; brief pause on the intro stats strip. | "We streamed two hundred thousand documents from FineWeb and measured them. Ninety-six snapshots of the web, seventy-thousand domains, twelve years. Here's what we found." | |
19 | | -| **0:22-0:42** | **Funnel.** Scroll step-by-step so stages reveal one at a time. Hover the **MinHash dedup** stage to pop the tooltip. | "It starts with the entire web: about a hundred trillion tokens of raw noise. Watch it get filtered: extract the text… keep only good English… and the biggest cut of all, remove the duplicates. What survives? Just fifteen trillion tokens. **Eighty-five percent is thrown away.**" | |
20 | | -| **0:42-0:56** | **Document length.** Land on linear view, then click **"Log scale"** yourself. | "How long is a typical web page? Tiny, about three hundred tokens. But flip to a log scale and the shape snaps into a clean bell. The web's documents are log-normal, with a few giant outliers carrying huge amounts of text." | |
21 | | -| **0:56-1:08** | **Domain galaxy.** Let bubbles settle; click a category swatch (e.g. **News**); type a domain in **search** (e.g. "wikipedia"). | "Where does all this text come from? Each bubble is a domain: news, wikis, blogs, code. You can filter by type, or search for any site… and there's Wikipedia, one of the giants." | |
22 | | -| **1:08-1:24** | **Concentration.** Show Lorenz curve + Gini; hover once; then click **"Token coverage."** | "But the web isn't fair. A handful of domains dominate, with a Gini of nearly point-six. In fact, just **three hundred domains** supply a quarter of *all* the text a model reads." | |
23 | | -| **1:24-1:38** | **Quality landscape.** Scroll so the dominant cell highlights. Hover the brightest cell. | "Cross length with language confidence and the whole dataset collapses into one bright region: confident English, medium length. That's the heart of what AI learns from." | |
24 | | -| **1:38-1:52** | Scroll to **takeaways** cards; slow pan across the four cards; end on the closing line. | "So the next time a model answers you, remember: it learned to speak from a filtered, deduplicated echo of the open web. We built this to make that invisible dataset something you can actually *see*, and explore yourself." | |
25 | | -| **1:52-2:00** | End card: title + live URL + "COM-480 EPFL". | *(silent, or)* "Decanting the Web. Thanks for watching." | |
| 19 | +| **0:22-0:42** | **Funnel**, revealed one stage at a time. Hover the **MinHash dedup** stage to pop the tooltip. | "It starts with the entire web: about a hundred trillion tokens of raw noise. Watch it get filtered: extract the text, keep only good English, and the biggest cut of all, remove the duplicates. What survives? Just fifteen trillion tokens. **Eighty-five percent is discarded.**" | |
| 20 | +| **0:42-0:56** | **Document length.** Land on the linear view, then click **"Log scale"**. | "How long is a typical web page? Tiny, about three hundred tokens. But flip to a log scale and the shape snaps into a clean bell. The web's documents are log-normal, with a few giant outliers carrying huge amounts of text." | |
| 21 | +| **0:56-1:08** | **Domain galaxy.** Let the bubbles settle; click a category swatch; search a domain (e.g. "wikipedia"). | "Where does all this text come from? Each bubble is a domain: news, wikis, blogs, code. You can filter by type, or search for any site, and there's Wikipedia, one of the giants." | |
| 22 | +| **1:08-1:24** | **Concentration.** Show the Lorenz curve and Gini, then click **"Token coverage."** | "But the web is not evenly distributed. A handful of domains dominate, with a Gini of nearly point six. Just **three hundred domains** supply a quarter of *all* the text a model reads." | |
| 23 | +| **1:24-1:38** | **Quality landscape.** Scroll so the dominant cell highlights. | "Cross length with language confidence and the whole dataset collapses into one bright region: confident English of medium length. That is the heart of what the model learns from." | |
| 24 | +| **1:38-1:52** | Scroll to the **takeaways** cards; end on the closing line. | "So the next time a model answers you, remember: it learned to speak from a filtered, deduplicated echo of the open web. We built this to make that invisible dataset something you can actually *see*, and explore yourself." | |
| 25 | +| **1:52-2:00** | End card: title, live URL, "COM-480 EPFL". | "Decanting the Web. Thanks for watching." | |
26 | 26 |
|
27 | 27 | --- |
28 | 28 |
|
29 | | -## 🎯 What to emphasize (contributions, not code) |
30 | | - |
31 | | -- **The funnel reveal** is the signature moment, so give it room. |
32 | | -- Show **interactivity you built**: the log toggle, the galaxy search/filter, the Lorenz↔coverage toggle. Each click should feel effortless. |
33 | | -- Land **one number per scene** so the viewer leaves with concrete facts (100T→15T, ~340 tokens, 300 domains, 92% ≥0.95). |
34 | | -- Tone: confident, curious, a little playful. You're a guide, not a lecturer. |
35 | | - |
36 | | -## ✂️ If you're over time |
37 | | -1. Merge the intro beat (0:10-0:22) into the hero. |
38 | | -2. Cut the quality-landscape beat to ~8s (it's the most technical). |
39 | | -3. Tighten the galaxy beat to a single search action. |
40 | | - |
41 | | -## 🎙️ Production checklist |
42 | | -- [ ] 1080p, 30fps, clean browser, full-screen site. |
43 | | -- [ ] Smooth slow scrolling (consider a scroll-automation bookmarklet). |
44 | | -- [ ] Narration recorded separately, normalized, light noise reduction. |
45 | | -- [ ] Soft background music bed at ~15% volume (optional). |
46 | | -- [ ] Final length **≤ 2:00**. Export H.264 MP4. |
47 | | -- [ ] Upload (YouTube unlisted / Drive) and paste the link into the README and process book. |
| 29 | +## Highlights |
| 30 | + |
| 31 | +- The **funnel reveal** is the signature moment of the piece. |
| 32 | +- Interactions demonstrated: the linear / log toggle on the length histogram, the domain-galaxy search and category filter, and the Lorenz / token-coverage toggle. |
| 33 | +- One concrete figure per scene: 100T to 15T tokens, a median near 340 tokens, about 300 domains for a quarter of all tokens, and 92% of documents scoring at least 0.95. |
0 commit comments