Lightnews — Scholar-powered news

bramvanroy.bsky.social @bramvanroy.bsky.social · 5d

Not sure if intentional or not, but it says "backTreasoning" in the dataset viewer screenshot slide.

1 1

bramvanroy.bsky.social @bramvanroy.bsky.social · 7d

Open training code and evaluation code could be added to get to fully open source.

1

bramvanroy.bsky.social @bramvanroy.bsky.social · 17d

Something that I've bumped my head against too often in `transformers` is how kwargs are sometimes dealt with. Just popping what's needed from kwargs but as a user it's then unclear if an argument is unexpectedly unused (and no good code editor support). Nice to see changes here!

1

bramvanroy.bsky.social @bramvanroy.bsky.social · 28d

Preparations going well!

1

bramvanroy.bsky.social @bramvanroy.bsky.social · Aug 22

Oh, and don't forget to register! ▶️ clin35.ccl.kuleuven.be/registration Hope to see you at CLIN in Leuven!

Computational Linguistics in the Netherlands @clin35-2025.bsky.social · Aug 22

🔊 #CLIN35 is three weeks away! The programme has been online for a while yet mystery remained around the keynote speaker but the wait is over! 🥳 Marie-Catherine de Marneffe will present "Consensus is a myth: Human label variation in Natural Language Inference". Info: clin35.ccl.kuleuven.be/programme

CLIN35 - Programme

The full programme with parallel sessions can be found here. The break-down for the poster sessions is here.

clin35.ccl.kuleuven.be

1

bramvanroy.bsky.social @bramvanroy.bsky.social · Aug 17

All great suggestions, thanks!

bramvanroy.bsky.social @bramvanroy.bsky.social · Aug 15

I'd like to grow my network more on this site as it is small compared to on that other site. My interests are quite broad so drop some names to follow 👇 around these topics: #machinetranslation #MT, Universal Dependencies, #LLMs in all of its aspects (technical, ethical, data), #Dutch NLP, #OSS!

1 2 4

Reposted

Computational Linguistics in the Netherlands @clin35-2025.bsky.social · Aug 8

🥳 Happy to open up the registrations for the CLIN conference! You can find more information here: clin35.ccl.kuleuven.be/registration The website has also been updated with more information for the presenters, with a programme, and with information about the venue. See you soon at #CLIN35!

1 2 3

Reposted

Computational Linguistics in the Netherlands @clin35-2025.bsky.social · Jul 9

🥳 We've sent out the notification of acceptance to all abstract submitters! Registration to CLIN will be opened up soon. We're really looking forward to seeing so many of you in Leuven on September 12th!!

1 2

bramvanroy.bsky.social @bramvanroy.bsky.social · Jul 2

Did you upload the paper? I would not trust these services when they say "they have read something" unless you explicitly provide the material to them.

1 1

Reposted

ctt2025.bsky.social @ctt2025.bsky.social · Jun 30

As a wrap-up for this year's CTT edition, we invite you to take a peek at our workshop's draft proceedings, in which you'll find the full papers of our wonderful speakers. The final version will be published ont he ACL Anthology, but in the meantime:
ctt2025.ccl.kuleuven.be/proceedings

CTT 2025 - Proceedings

The draft proceedings are available at this link: https://drive.google.com/file/d/1yCOHAHTpqsyFCL2c5jj7Khy6yyKJ7BbC/view?usp=drive_link Important note: these are just the drafts, provided for your vie...

ctt2025.ccl.kuleuven.be

2 7

bramvanroy.bsky.social @bramvanroy.bsky.social · Jun 25

Really cool! Going to forward this to some people. Maybe you can add a line on how the "📜 Traditional OCR Output" is generated? Or is that what's in the optional XML?

1 1

bramvanroy.bsky.social @bramvanroy.bsky.social · Jun 25

I could not be there this year and mostly worked behind the scenes, but I am very happy to see the workshop growing! Hope everyone had a nice time!!

ctt2025.bsky.social @ctt2025.bsky.social · Jun 25

The #CTT2025 workshop ended yesterday after an insightful day into new creative domains, new applications and tools, new projects and new avenues for research and action. Many thanks, once again, to our presenters, and to everyone who participated! Now, on to the Summit itself!

1 2

Reposted

ctt2025.bsky.social @ctt2025.bsky.social · Jun 24

#CTT2025 is now live at MT Summit 2025!

3 6

Reposted

Stella Biderman @stellaathena.bsky.social · Jun 7

EleutherAI is trying to build bridges with libraries and other cultural heritage institutions, but I would also love to talk with digital humanities students or researchers about how we can collaborate in mutually beneficial ways.

3 5 23

bramvanroy.bsky.social @bramvanroy.bsky.social · Jun 2

Yay, so happy to host CLIN in Leuven this year! It'll take place on September 12th. Abstract submission deadline on June 13th!

Computational Linguistics in the Netherlands @clin35-2025.bsky.social · Jun 2

📅 Don't forget! The deadline for submitting your abstract to the #CLIN conference in Leuven is coming: 13th of June! Submitting is easy: name, title of your work, 500-word abstract, done! #nlp #nlproc #compling #llm #ai #dutch clin35.ccl.kuleuven.be

CLIN35

Computational Linguistics in The Netherlands (CLIN) is a yearly conference on computational linguistics. Each year the conference is organized by a different institution in the Dutch-speaking region. ...

clin35.ccl.kuleuven.be

1 3

bramvanroy.bsky.social @bramvanroy.bsky.social · May 4

More work needs to be done! Only 7 out of 100+ crawls have been processed so far. That's encouraging because it means there is a lot more Creative Commons data to be collected! If you have the compute available and wish to collaborate in an open and transparent manner, please get in touch!

1

bramvanroy.bsky.social @bramvanroy.bsky.social · May 4

In total, this V1 release contains almost 150 thousand documents and 150 billion tokens. This data was not filtered on quality nor deduplicated so that you can decide for yourself how much data to keep or discard but a dataset field describes whether a doc is included in the FineWeb(-2) datasets.

1 1

bramvanroy.bsky.social @bramvanroy.bsky.social · May 4

In this first version of C5, 8 languages are included (afr, deu, eng, fra, fry, ita, nld, spa). The language set was limited for 2 reasons: computational limitations and a collaboration with GPT-NL, who requested CC data for these languages to train a Dutch-focused, copyright-conscious #LLM.

1

bramvanroy.bsky.social @bramvanroy.bsky.social · May 4

To build C5 HTML pages are scrutinized and links to CC licenses are collected. Additional data fields are included such as "was the license found in the HTML `head`?" or "if multiple licenses were found, do they contradict each other?", which makes further filtering a breeze.

1

bramvanroy.bsky.social @bramvanroy.bsky.social · May 4

📢 Introducing the Common Crawl Creative Commons Corpus (C5)!

An effort to filter @commoncrawl.bsky.social web-crawls to only documents that are @creativecommons.bsky.social licensed. At this stage 150 billion tokens have been collected - more to come!

More info here: huggingface.co/posts/BramVa...

@BramVanroy on Hugging Face: "📢💾 Introducing the Common Crawl Creative Commons Corpus (C5)! C5 is a…"

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co

1 2

Reposted

AT4SSL Workshop @at4ssl.bsky.social · Mar 18

💡 We extend the submission deadline for AT4SSL! If you were working on a submission but didn't manage to finish it we can now offer a few days more so you can still submit it. The new deadline is ⏰ Thursday 20/3 AOE ⏰. All information at sites.google.com/tilburgunive.... @mtsummit2025.bsky.social

Home

SCOPE The rapid technological and methodological advances in deep learning, and in AI in general, that we see in the last decade, have not only improved MT, recognition of image, video and audio signa...

sites.google.com

1 3 4

bramvanroy.bsky.social @bramvanroy.bsky.social · Mar 11

Oooh, the LLM world definitely has A LOT to learn when it comes to evaluation best practices. Looking forward to seeing WMT diligence and years of experience and improvement being applied in this field too.

5

bramvanroy.bsky.social @bramvanroy.bsky.social · Mar 7

Isn't implementing task routes an architectural change over vanilla transformer? Basically task-specific heads, potentially with some MoE in there for routing?

1 1

bramvanroy.bsky.social @bramvanroy.bsky.social · Mar 7

Agreed. I also think that in the future the tasks (auto-regression vs diffusion vs pure token-level prediction) will somehow merge into an architecture that can do whatever seems more optimal. At least I imagine something has to happen architecturally to counter scaling-ceiling in an efficient way.

1 1