Yoel Zimmermann

Scalable watermarking for identifying large language model outputs - Nature A scheme for watermarking the text generated by large language models shows high text quality preservation and detection accuracy and low latency, and is feasible in large-scale-production settings.

www.nature.com/articles/s41...

What there should be is a *single* place where one deposits a standard letter for all students. This would be much more scalable and sustainable.

YESSS!!!

P.S. If you work in the flavor/food/fragrance industry and are laughing about our 15k dataset, reach out to hello@kvantlabs.ai. We would love to build an even better model for you!

A Chemical Language Model for Molecular Taste Prediction Determining molecular taste remains a significant challenge in food science. Here, we present FART (Flavor Analysis and Recognition Transformer), a chemical language model capable of predicting molecu...

Preprint: chemrxiv.org/engage/chemr...

Dataset: huggingface.co/datasets/Far...

It has been fun working on this with @leif7ieben.bsky.social, @heseng.bsky.social, Philipp Pestlin, Franz Görlich!

Trained on the largest publicly available dataset (15,025 unique compounds), we achieve >90% accuracy - outperforming existing binary classifiers that focus on only one taste class (e.g. sweet/non-sweet).

Try it out here: fartlabs-fart.hf.space

Ever had the sudden urge to taste something in the lab?

Now you can! (For legal reasons: you can't)

FART (Flavor Analysis and Recognition Transformer) is the first machine learning model to simultaneously predict four taste classes – sweet, sour, bitter, umami - from chemical structure alone!

What a time to be alive

A screenshot of overleaf.com giving a "Error: Server Error" message.

Me, a moron: I don't need a local TeX installation, Overleaf works so well and is so easy!

Overleaf: like every online service that exists, goes down from time to time.

Me: shocked Pikachu.

Is 15 minutes supposed to be a lot?

CERN also came to my mind 😅

For those of you looking for a big helping of LLMs for Thanksgiving, here you go! 🤖🦃

📝 arxiv.org/abs/2411.15221

We release 34 examples demonstrating applications of LLMs in materials science and chemistry across:
🔶molecular and material property prediction;
🔷molecular and material design;
...

In principle you can do a static evaluation on a holdout set but I have no clue if 4k datapoints is enough.

Cool work! How about doing RLHF on your data?

If only scholarly publications came with a short synopsis of the paper right up front, written by the authors to highlight the important and salient points of the study.

We could give it a nifty name, like "abstract".