Want to explore #wordembedding properties? Here are some ideas for data preparation steps✂️, interesting additional features🍩, and some relevant references✍️: embedding-framework.lingvis.io With
Menna El-Assady 😊 #nlviz #ieeevis #LLMs

Our visualizations show that larger models, e.g., ChatGPT and Bard, are less biased than their smaller counterparts. Are the models getting better or the engineering steps used to adapt the outputs? 🤔Menna El-Assady #ieeevis #vds 👉prompt-comparison.lingvis.io

What biases are encoded in texts generated by #LLMs ? Our workspace helps to explore stereotypes encoded in prompt outputs and detect unexpected word associations. Menna El-Assady will present our work
VDS at KDD and VIS🥳 Paper and demo: prompt-comparison.lingvis.io #ieeevis2023 IEEE VIS

Ian Johnson 💻🔥 Rita Sevastjanova Thanks a lot for your effort! 😊
I also figured that would be the correct CLIP version. However, regardless if I try the laion/CLIP-ViT-H-14-laion2B-s32B-b79K model from HuggingFace or your linked code with the open-clip-torch pip package, I get embeddings differing from SDv2.1.

Featuring some great work from Steffen Eckhard, Vytautas Jankauskas, Elena Leuschner, @MisterXY89 & Rita Sevastjanova in this latest post, with helpful tips for practitioners on harnessing AI.

Yes! Hallucination in LLMs is not a bug👇

Can someone help us out on this? #stablediffusion

Thilo Spinner Rita Sevastjanova yeah, i forgot where i found the relevant line, but i made this little server to give me both the old clip embeddings for 1.x and new ones for 2.x
gist.github.com/enjalot/222895…

i made a 'full_encode' method that borrowed an extra piece from the transformers lib i think

Ian Johnson 💻🔥 Rita Sevastjanova Okay, by chance, I found that SDv2.1 uses the penultimate (second to last) layer. By that, I get the correct embeddings. 🎉
Thanks a lot for your hints - without experimenting with the FrozenOpenCLIPEmbedder I would not have noticed this! 😌

Does this procedure work? We think so! Check out our validation of the method and data here: doi.org/10.1007/s11558…

Rita Sevastjanova Ian Burton

Menna El-Assady Congrats!!! 🎊

Ian Johnson 💻🔥 Rita Sevastjanova I feel like I'm missing something here...
Did you compare the embeddings between CLIP-ViT-H-14-laion2B-s32B-b79K and the SDv2.1 text encoder?

There are so many people to thank! This has been so long in the making and we hope we added a complete list in the paper 😊

A not complete list + only ppl on Twitter: Ian Burton, Eva Thomann @evathomann.bsky.social, Ronny Patz, Mirko Heinzel, Laurin Friedrich, @kateweaverUT, Rita Sevastjanova

Brendan Dolan-Gavitt @idkmyna40085148 If it is still relevant: Roberta (encoder-decoder) fine tuned for code classification task (including C++): huggingface.co/NTUYG/DeepSCC-… I would take the CLS token (here: <s>) embeddings from the second to last layer.

PEIO/RIO Uni Konstanz Department of Politics & Public Admіn Thank you Steffen Eckhard, Vytautas Jankauskas, Ian Burton, Rita Sevastjanova, and Tilman Kerl for the amazing time working together for the past years!

Ian Johnson 💻🔥 Thilo Spinner Thanks, Ian Johnson 💻🔥, for your help🙂

Brendan Dolan-Gavitt @idkmyna40085148 Yes. But I think the problem here is the input data - GPT2 has never seen code during its training - rather than the model’s type.

Thanks to the stellar team Vytautas Jankauskas Elena Leuschner Rita Sevastjanova, I.Burton + T.Kerl and zeppelin universität Universität Konstanz DFG public | @[email protected] for the support!

Karsten Donnay PEIO/RIO Vytautas Jankauskas Rita Sevastjanova Elena Leuschner Thanks Karsten. This one has been a lot of fun, too. Very lucky when it turnes out well on top 😀