Hanna Yukhymenko

post-training apertus @eth_ai_center | ex-🤗 @eth | EEU ambassador @insaitinstitute

[Hanna Yukhymenko @a_yukh]

Hot open-source summer season has gotten to Switzerland 🇨🇭⛰️

Our team at Swiss AI Initiative is happy to release Apertus 1.5 - a multimodal and reasoning update to our first v1 version!

The new version builds a foundation for future regular open model development at scale🧵

[Hanna Yukhymenko @a_yukh]

@yacinelearning finally a very detailed podcast on current state of evals, been looking for that last week literally

[Hanna Yukhymenko @a_yukh]

so many nice benchmarks cooking rn, maybe I am an eval engineer after all

[Hugging Models @HuggingModels]

Check out Apertus-v1.5-70B.

[Hanna Yukhymenko @a_yukh]

Not only reports are extremely vague, but also there is very few data for safety training, only some from Nemotron.
All red-teaming competitions have low payouts and don't open source prompts either; for highly paid area closed companies get this data basically for free.

[Florian Brand @xeophon]

@tenobrus The thing that is jarring to me is how closed the safety ecosystem is.

A lot of the reports are cryptic, software mostly isn’t as open as other parts of the stack and methodologies (like aligning GPT-OSS) are kept secret.

[Hanna Yukhymenko @a_yukh]

@Laz4rz I think it's missing the word "Eastern"

[Hanna Yukhymenko @a_yukh]

@adidshaft @apertusllm gating on HF is a legal requirement in order to help make sure that the usage terms and other policies are complied with, currently we are working on clarifying this

[Hanna Yukhymenko @a_yukh]

Hot open-source summer season has gotten to Switzerland 🇨🇭⛰️

Our team at Swiss AI Initiative is happy to release Apertus 1.5 - a multimodal and reasoning update to our first v1 version!

The new version builds a foundation for future regular open model development at scale🧵

[Hanna Yukhymenko @a_yukh]

@Laz4rz Polish AGI benchmark

[Hanna Yukhymenko @a_yukh]

@Shekswess @apertusllm Yesss since last December! Hot OSS summer indeed 🙌

[Hanna Yukhymenko @a_yukh]

@Laz4rz @xeophon brachu 🥺🥺

[Hanna Yukhymenko @a_yukh]

@Laz4rz unc is NOT losing it

[Hanna Yukhymenko @a_yukh]

We are working hard on preparing all training artifacts including technical report very soon! (checkpoints, data, more evals too - stay tuned)

For now you can test our models through HF and PublicAI https://t.co/Ki7BZW0Vyq
8B: https://t.co/dbXQF8ziJZ
70B: https://t.co/AHZrMIef4J

[Hanna Yukhymenko @a_yukh]

On the text side, we have focused on improving instruction following and tool calling together with introducing reasoning capabilities. New 1.5 version shows great progress compared to v1 and catches up with newer open-source and open-weight alternatives.

[Hanna Yukhymenko @a_yukh]

@Laz4rz @arcee_ai Bros cooking

[Hanna Yukhymenko @a_yukh]

@Laz4rz Can confirm, miles better

[Hanna Yukhymenko @a_yukh]

Tired of seeing the same crazy numbers on WMT or FLORES but the model is still bad in your language irl?
Join us in designing a new translation benchmark to create a real challenge for language models👀

[Vilém Zouhar @zouharvi]

There are many things machine translation still can't do. Help us steer the next direction by contributing hard-to-translate inputs (and be on a cool paper).

[Hanna Yukhymenko @a_yukh]

50 stages of safety training:

[Hanna Yukhymenko @a_yukh]

I step foot in LA and I see this instantly what @tbpn