Yoni Gottesman

avoid side effects

[Yoni Gottesman @yonigottesman]

tiny macos app that fixes grammar and typos. perfect for quick messaging ⌨️
all local using gemma-4-E2B-it-qat-GGUF

https://t.co/h8Wavg7b5y

[Yoni Gottesman @yonigottesman]

had some fun playing with late interaction visual embedding models for medical text and wrote about it.
tl;dr nvidia/nemotron-colembed-vl-8b-v2 is a good model :)
https://t.co/rfFD8dYO1S

[Yoni Gottesman @yonigottesman]

also works with images

[Yoni Gottesman @yonigottesman]

I made a tiny tool to manually eyeball your retrieval results.
Just give it your search function with any parameters and it will generate a simple ui.
https://t.co/P3GBy4ne3Q

[Yoni Gottesman @yonigottesman]

I made a minimal agent to explore Mimic-III dataset and wrote about it.
You don't need any fancy agent library to do cool things ;)

https://t.co/wwhkmk27qa

[Yoni Gottesman @yonigottesman]

@shaiyallin קריאה מעולה, תודה!
מפרמל בצורה יפה המון מחשבות שלי, ויהיה לי עכשיו יותר קל לדברר :)

[Yoni Gottesman @yonigottesman]

Fun experiment: Training a custom model using TotalSegmentator-generated weak labels, powered by Monai + Keras 3 w/ PyTorch backend
https://t.co/Lkq1gNdDqY

[Yoni Gottesman @yonigottesman]

https://t.co/9RuVZwDKAd

[Yoni Gottesman @yonigottesman]

I've added a new feature to @huggingface `apply_chat_template` that makes it return a mask of assistant-generated tokens, making it easier to label other tokens with -100 during training. Previously, I had to manually tokenize the conversation and keep track of these tokens. 🥳

[Yoni Gottesman @yonigottesman]

@yuntiandeng and linked to the exact function you did 😎

[Yoni Gottesman @yonigottesman]

I think most finetuning scripts are sub-optimal in that they do not mask user parts of the conversation.
this includes popular training scripts:
* hf alignment-handbook
* mlabonne llm-course
* hf autotrain
I did some experiments and wrote about it here

https://t.co/Oh3y7MwqVI

[Yoni Gottesman @yonigottesman]

@yuntiandeng I did mention fastchat for this in my post! ;-)

[Yoni Gottesman @yonigottesman]

i suggested a feature in the apply_chat_template abstraction that will make it return the labels with -100 on user tokens. https://t.co/9RuVZwEipL this can make training scripts more aligned with inference code

[Yoni Gottesman @yonigottesman]

@BramVanroy the loss in both cases is computed the same, only on assistant tokens. BUT of course you are right! this should further be tested on some downstream task.

[Yoni Gottesman @yonigottesman]

@IgorMintz אני אבא ואני מאשר את הבדיחה הזאת

[will depue @willdepue]

I ask DALLE-3 to generate a Pepe but each time I tell it to make it "more rare."

[Yoni Gottesman @yonigottesman]

really proud I showed a nice example without wrapping it in a package like everyone else :)

[Yoni Gottesman @yonigottesman]

I wrote how to do "Extractive QA" with a decoder model like llama. This technique guarantees the generated output will 100% be a substring of the context.

https://t.co/mQPQaWn88E

[Yoni Gottesman @yonigottesman]

@IgorMintz

[Yoni Gottesman @yonigottesman]

@DaphnaShalevWG @IgorMintz @IncubatorTheatr מה זה לא ידעתי שיש פיד וויז