serhii.net

Fought long and hard on this till I got enlightened. The trivial stupid way works.

\providecommand{\dagtab}{%
{\textsuperscript{\dag}}
}
\providecommand{\asttab}{%
{\textsuperscript{*}}
}

% put it wherever in the table

\caption[Evaluation scores]{Scores of selected models. 

\dagtab LMES tasks (shortened for brevity)

This was so easy.

\ddag double dagger also exists
! as well

The UNLP workshop generously included a 3 months trial of Grammarly Premium, and this was interesting.

• Shown in Fig. XXX

• Oxford commas — Grammarly wants them, I seem to not, but I should decide on one

• it specified the requirements, THE complexity of the story, … — when I list things I can’t use one article for all of them!

  • I would never write “I have a cat, dog, rabbit”, and that’s the same thing

• e.g. requires a comma only in American English

• I do A LOT of errors with duplicated words (the the)

  • A book I’m reading suggests using a different font and column width when proofreading just to catch such things

• I don’t know how to spell a number of English words

  • accommodate

• Many typos are acoustical ones if I’m tired, e.g. doc instead of dog

• I use too many words

  • A lot of my passive can be really neatly rephrased into not-passive, for shorter sentences with fewer words
  • verb+noun often can be just a better verb (like make predictions => predict)

• it removes many of my commas

  • I know it’s my Slavic language interference, but in many cases it makes the text way less clear imo

General impressions

• It even knows that written in python requires Python to be capitalized!

Comma after ‚i.e.‘ and ‚e.g.‘ – Business English & Übersetzungen:

• i.e. means that is / in other words (from id est)
• e.g. means for example

They are followed by commas in American English but not in British English.

kubectl cp failed with errors, so.

file.io - Super simple file sharing

(rapids) root@lm-eval-sh:/data/output# curl -F "file=@more.zip" https://file.io

{"success":true,"status":200,"id":"xxx","key":"xxx","path":"/","nodeType":"file","name":"more.zip","title":null,"description":null,"size":46277219,"link":"https://file.io/xxx","private":false,"expires":"xxx","downloads":0,"maxDownloads":1,"autoDelete":true,"planId":0,"screeningStatus":"pending","mimeType":"application/octet-stream","created":"2024-04-16T15:19:10.227Z","modified":"2024-04-16T15:19:10.227Z"}

Generally, free curl file sharing online - Google Suche returns many services with potential.

seaborn.barplot — seaborn 0.13.2 documentation:

passing order=[list,of,cats,in,order] decides the ordering.

Otherwise “it will be inferred” except that it’s not always trivial to understand how exactly (or I’m too sleep-deprived).

And if I’m drawing horizontal lines on top of the bars in the barplot based on indexes then the order may be sligthly different.

With the help of ChatGPT

\documentclass{article}
\usepackage{titlecaps}
\usepackage{etoolbox}

% Specify words to remain in lowercase unless they are the first word
\Addlcwords{the and but or nor for a an at by to in on with of}

\let\oldchapter\chapter
\renewcommand{\chapter}[1]{\oldchapter{\titlecap{#1}}}

\let\oldsection\section
\renewcommand{\section}[1]{\oldsection{\titlecap{#1}}}

\let\oldsubsection\subsection
\renewcommand{\subsection}[1]{\oldsubsection{\titlecap{#1}}}

\begin{document}

\section{an example of a section with and without uppercasing specific words}
This is some text.

\subsection{exploring the integration of tools in the workplace}
More text here.

\end{document}

The same thing can have multiple names and that’s alright!

\label{old-subsection-name-maybe-linked-to-elsewhere}
\label{sec:eval-task-2}

Margin notes - Overleaf, Online-LaTeX-Editor:

\marginpar{text} is the vanilla option, but this works in all cases ever:

\usepackage{marginnote}
\marginnote{text}

EXCEPT I couldn’t find a way to add footnote markers to have numbered margin notes separate from the real footnotes.

But this solves everything, quoting directly¹:

\newcounter{mgncount}
\renewcommand\themgncount{\arabic{mgncount} }
\newcommand\marginfootnote[1]{\refstepcounter{mgncount}\marginpar{{$^\themgncount$}#1}\footnotemark}

\begin{document}
Can we put a footnote with number in the margin and a number in the text?\marginfootnote{There's a number here!}

Another test\marginfootnote{Working!}

One more try\marginfootnote{Successful!}
\end{document}

EDIT: actually it doesn’t and uses the number from footnotes in the text itself. :(

Ah, the sidenotes package exists: https://ctan.math.utah.edu/ctan/tex-archive/macros/latex/contrib/sidenotes/sidenotes.pdf

But uses only 1…3-type numbers.

Yes this is it! CTAN: Paket sidenotesplus

%\usepackage{sidenotes}
\usepackage[mark=Alph]{sidenotesplus}
...

\sidenote{does basically what footnote does}

It has a lot of options and can do a lot of things, yes, this is it, it’s perfect. The example page has everything: https://ctan.math.illinois.edu/macros/latex/contrib/sidenotesplus/tests-sidenoteplus.pdf

See also CTAN: Marginal topic.

marginpar - Footnote and number in margin - TeX - LaTeX Stack Exchange

symbols - What is the best way to use quotation mark glyphs? - TeX - LaTeX Stack Exchange:

``this'' / `this' is the proper way
"this"/'this' produces two closing quotes and 'is annoying to readers'

There’s also CTAN: Package csquotes that ‘is fantastic’, including smartly doing nested quotes, correct quotes for diff languages, and ‘generally always doing what you want it it’:

\usepackage[autostyle]{csquotes}

...
\enquote{My quoted text}

Another answer suggests

\newcommand{\q}[1]{``#1''}

...

\q{whatever}

I’ve been using more custom latex commands lately and this goes in that direction.

I guess creating a \q that does autoquotes w/ csquotes is the way to go?

• Punctuation

  • all citations to word~/cite{xxx}.
  • all footnotes to sentence.\footnote{}¹
    • either full sentence or lowercase part
    • but within parentheses!
  • for both, it’s sent~\cite{}.\footnote{}
  • all numbers to 132,32.99
  • Consistent quoting (using the correct latex quotes or \enquote{} with italics for longer sentences.)
  • all refs to autorefs
    • Autoref fails with appendix subsections, do it manually.
  • tightlists everywhere
  • Overleaf ‘stop on first error’ to fix the errors
  • Title Case in all Titles

• Bits

  • CBT-UA -> UA-CBT
  • LMentry-static-UA shorten to LMES once and keep using LMES.
  • Eval-UA-tion should be captialized
  • Thesis always capitalized
  • gpt2/GPT2 -> GPT-2.
  • check for stray ‘we’s in the paper
    • “our”/“we” “paper”
  • look for sticking out over-the-line bits
  • Python is capitalized
  • all Grammarly suggestions

• Not bits

  • go through all latex comments
  • go through all latex warnings
  • go through all todos in home.md + taskwarrior

====== Open research questions:

• Research
  • look into whether translated datasets are worse at stuff
  • monolingual VS multilingual models incl Ukrainian performance
  • Whether prompt language makes a differenec on Ukrainian task
• Datasets:

SH, [10 Apr 2024 14:58:39] LMES — дослідити robustness моделей, і наприклад глянути яка залежність accuracy людей і ШІ в залежності від мммм різниці в довжині слів чи номеру слова (“яке стотринадцяте слово в реченні …”) CBT-UA — нормально evaluate, а ще для людей і машин — глянути scores якщо давати тільки challenge segment. Я це тестив з нейромережами (не попало в paper), але там дуже часто були кращі результати з фрагментом ніж з усією казкою

SH, [10 Apr 2024 14:59:57] Зробити датасет по biases і фемінітивам, у мене написаний код для генерації нульової версії, там по суті речення типу “моя жінка займається програмуванням компʼютерних систем, тобто за професією вона — ….”

SH, [10 Apr 2024 15:00:20] Мрія всього життя таки зробити Russian-Ukrainian interference dataset на предмет русизмів та російських помилок

SH, [10 Apr 2024 15:02:57] UA-CBT — взяти казки з project Gutenberg, взяти іноземні казки перекладені українською, і порівняти scores моделей на тасках по казкам з цих різних джерел. Можна забити на фільтрацію, чисто зробити human baseline на частині згенерованого датасету. Так можна зробити нереально великий датасет і знати що там максимум умовнио 80% бо 20% тасків сміття

Also:

• CATSMC and friends — much larger datasets can be generated from the given data, a lot of combinations are possible.