Spaces:

---

CYFRAGOVPL

/

README

like 2

Running

If you really want an answer, it's best to write to https://pllum.org.pl/#contact. It's unlikely anyone will answer you here.

To be quite honest this model seems like a grant scam, there's no model card, no research paper, zero information about training, zero benchmark results.
There are only sparse inforrmation on where it was trained, that it apparently 'achieved state-of-the-art performance in internal benchmarks', and what models are the base ones.
It's sad to see a model like this being promoted by our national goverment while a better Polish LLM made in free time gets no such support or funds,.
https://huggingface.co/speakleash/Bielik-11B-v2.2-Instruct

To be quite honest this model seems like a grant scam, there's no model card, no research paper, zero information about training.
The only information that's out there is that it cost 15 million PLN and that the chat model was trained on Mistral 8x7B(a model from December 2023).
It's sad to see a model like this being promoted by our national goverment while a better Polish LLM made in free time gets no such support or funds,.
https://huggingface.co/speakleash/Bielik-11B-v2.2-Instruct

Bielik is an amateur attempt, PLLuM beats it in every way, a lot could have been refined, but it is so far the only real Polish LLM, not a taken Mistral and fine-tuned on a sparse dataset. It's a success for Krzysztof Gawkowski, the documentation etc. is not transparent yet, but if you have the right contacts you can get additional information (this doesn't mean it doesn't exist). It's just that PLLuM is not commercial and its capabilities are narrow, so publishing a preprint etc. can wait. I have written many articles on this topic

To be quite honest this model seems like a grant scam, there's no model card, no research paper, zero information about training.
The only information that's out there is that it cost 15 million PLN and that the chat model was trained on Mistral 8x7B(a model from December 2023).
It's sad to see a model like this being promoted by our national goverment while a better Polish LLM made in free time gets no such support or funds,.
https://huggingface.co/speakleash/Bielik-11B-v2.2-Instruct

Bielik is an amateur attempt, PLLuM beats it in every way, a lot could have been refined, but it is so far the only real Polish LLM, not a taken Mistral and fine-tuned on a sparse dataset. It's a success for Krzysztof Gawkowski, the documentation etc. is not transparent yet, but if you have the right contacts you can get additional information (this doesn't mean it doesn't exist). It's just that PLLuM is not commercial and its capabilities are narrow, so publishing a preprint etc. can wait. I have written many articles on this topic

Yes it's an amateur attempt but only because of a lack of funding. The team works part-time and has real other jobs.
But they somehow still manage to be more transparent, created an actual benchmark, their own training framework, have a miles better site and most importantly don't make empty promises.
I can say that I created an LLM that beats even OpenAI o3 and has 100% in every benchmark, but these words hold the same value as the PLLuM promises that it beats their 'Polish-language tasks benchmarks.'
Like they say Talk is cheap, show me the work.
I don't necessarily think PLLuM is a bad model, but the lack of transparency is concerning, especially when it was annouced as a "first, open, transparent and strongly Polish-language generative model", but delivered less quality and transparency than an amateur attempt.

@Kajaqq If you have such a model, share it with us

I am struggling in starting and running the model, whilst all prerequisites are met.
At least all prerequisites mentioned in the card.
So, can someone respond to my simple question over the other thread, so that I can test it end to end?

Is it true the cost was around 15 million PLN for this project?
I heard there are some plans to integrate this with mObywatel and some documents processing tasks.
My current results are far away from acceptable results and I am wondering what did I do wrong with my setup (or is it the quality of project, not my setup)?

Where can I get some answers? Thanks.

To be quite honest this model seems like a grant scam, there's no model card, no research paper, zero information about training.
The only information that's out there is that it cost 15 million PLN and that the chat model was trained on Mistral 8x7B(a model from December 2023).
It's sad to see a model like this being promoted by our national goverment while a better Polish LLM made in free time gets no such support or funds,.
https://huggingface.co/speakleash/Bielik-11B-v2.2-Instruct

Bielik is an amateur attempt, PLLuM beats it in every way, a lot could have been refined, but it is so far the only real Polish LLM, not a taken Mistral and fine-tuned on a sparse dataset. It's a success for Krzysztof Gawkowski, the documentation etc. is not transparent yet, but if you have the right contacts you can get additional information (this doesn't mean it doesn't exist). It's just that PLLuM is not commercial and its capabilities are narrow, so publishing a preprint etc. can wait. I have written many articles on this topic

Please, I beg you, let’s not turn the AI community into politics! PLLuM and Bielik are two fantastic teams. This is not a scam, and they are not amateurs. I highly recommend the event at the Faculty of Mathematics and Computer Science at UAM – the whole truth about Polish LLMs. Take your time, explore, and watch in peace. Let’s not divide the Polish AI scene. Scientists, engineers, researchers, programmers, and data scientists have built two ecosystems for model development – a phenomenon on a European scale!

To be quite honest this model seems like a grant scam, there's no model card, no research paper, zero information about training.
The only information that's out there is that it cost 15 million PLN and that the chat model was trained on Mistral 8x7B(a model from December 2023).
It's sad to see a model like this being promoted by our national goverment while a better Polish LLM made in free time gets no such support or funds,.
https://huggingface.co/speakleash/Bielik-11B-v2.2-Instruct

Bielik is an amateur attempt, PLLuM beats it in every way, a lot could have been refined, but it is so far the only real Polish LLM, not a taken Mistral and fine-tuned on a sparse dataset. It's a success for Krzysztof Gawkowski, the documentation etc. is not transparent yet, but if you have the right contacts you can get additional information (this doesn't mean it doesn't exist). It's just that PLLuM is not commercial and its capabilities are narrow, so publishing a preprint etc. can wait. I have written many articles on this topic

Please, I beg you, let’s not turn the AI community into politics! PLLuM and Bielik are two fantastic teams. This is not a scam, and they are not amateurs. I highly recommend the event at the Faculty of Mathematics and Computer Science at UAM – the whole truth about Polish LLMs. Take your time, explore, and watch in peace. Let’s not divide the Polish AI scene. Scientists, engineers, researchers, programmers, and data scientists have built two ecosystems for model development – a phenomenon on a European scale!

Please stop humiliating yourself. The creator or co-creator of Bielik got so upset (like little children) after a discussion with a 20-year-old on Twitter that they started searching for me and disliking my comments, as well as finding my articles on Wikipedia. They want to build a serious and grand "European LLM model" (which doesn't reflect reality anyway) and they're all red-faced after a discussion with a 20-year-old on Twitter who criticized their project lol. They better be glad they got a computing unit from Cyfronet for training...

To be quite honest this model seems like a grant scam, there's no model card, no research paper, zero information about training.
The only information that's out there is that it cost 15 million PLN and that the chat model was trained on Mistral 8x7B(a model from December 2023).
It's sad to see a model like this being promoted by our national goverment while a better Polish LLM made in free time gets no such support or funds,.
https://huggingface.co/speakleash/Bielik-11B-v2.2-Instruct

Bielik is an amateur attempt, PLLuM beats it in every way, a lot could have been refined, but it is so far the only real Polish LLM, not a taken Mistral and fine-tuned on a sparse dataset. It's a success for Krzysztof Gawkowski, the documentation etc. is not transparent yet, but if you have the right contacts you can get additional information (this doesn't mean it doesn't exist). It's just that PLLuM is not commercial and its capabilities are narrow, so publishing a preprint etc. can wait. I have written many articles on this topic

Please, I beg you, let’s not turn the AI community into politics! PLLuM and Bielik are two fantastic teams. This is not a scam, and they are not amateurs. I highly recommend the event at the Faculty of Mathematics and Computer Science at UAM – the whole truth about Polish LLMs. Take your time, explore, and watch in peace. Let’s not divide the Polish AI scene. Scientists, engineers, researchers, programmers, and data scientists have built two ecosystems for model development – a phenomenon on a European scale!

Please stop humiliating yourself. The creator or co-creator of Bielik got so upset (like little children) after a discussion with a 20-year-old on Twitter that they started searching for me and disliking my comments, as well as finding my articles on Wikipedia. They want to build a serious and grand "European LLM model" (which doesn't reflect reality anyway) and they're all red-faced after a discussion with a 20-year-old on Twitter who criticized their project lol. They better be glad they got a computing unit from Cyfronet for training...

Please read the comments on Twitter/X :-) We try to keep the discussion fair and based on facts. We share research papers, frameworks, and techniques. No answers :-( If a young person makes strong claims, we try to talk about them, not argue. This is a discussion, not an attack. So again, I kindly invite you to watch: https://www.youtube.com/playlist?list=PLHqGfF79LBPXn09bptxc0R-fzGw-O-HXG
In my last post, I wanted to highlight two fantastic teams! I didn’t want to divide the Polish scene, but you responded with 'Please stop humiliating yourself.' ;-) So once again - peace. Let’s talk and avoid words like: 'amateurs,' 'scam,' 'mistake,' 'forget it,' or 'mere fine-tuning.' Let’s have a real discussion!