File size: 2,119 Bytes
077605d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
# Transformer attention-head ablation briefing

In the supplied `input/workspace/`, create **eight Beamer TEX slides plus PDF** in English for researchers at a group meeting. Preserve product names, paper titles, and necessary terminology.

Read `input/workspace/README.md` and `input/workspace/reading-scope.md` first. Originals are read-only; use scratch/ for intermediates and artifact/ for deliverables. Do not introduce unsupported facts. Original PDFs/webpages take precedence over conflicting auxiliary TXT/CSV files.

## Content requirements

1. The base has 8 heads and dev BLEU 25.8; 1/4/16/32 heads correspond to 24.9/25.5/25.8/25.4.
2. Identify the metric as newstest2013 development-set BLEU, not test BLEU or accuracy.
3. Explain that head dimensions also change in this experiment; do not claim all other parameter values remain fixed. Retain the Metropolis style.

## Deliverables

- Follow the format, page count, dimensions, and duration specified above; slide counts include the cover and conclusion.
- Use clear hierarchy and consistent typography, spacing, and colors. Avoid text overflow; keep graphics and citations readable.
- Preserve editable text, shapes, and data charts; a full-slide raster image must not be the sole content.
- Supply per-slide PNG previews with PPTX; supply compilable source, required assets, and PDF with TEX. PNG-only deliverables do not additionally require source files.
- Cite facts and images in the artifact or `sources.md`; explain final file locations and any necessary opening instructions in `delivery.md`.

Submit to `artifact/`. Choose your own visual layout; reproducing the source webpage or paper layout is not required.

## Scenario and presentation preferences

Scenario: academic presentation. Audience/channel: researchers in the relevant field, projected at a group meeting.

- Connect the ablation question, setup, results, and conclusion without confusing variables.
- Make head counts and metrics comparable, with readable table headings, units, and conditions.

Choose your own layout; reproducing a particular reference answer is not required.