No, the model has nothing do to with Llama. We are using our own architecture, and training from scratch. Llama also does not have open training data, and is non-compliant, in contrast to this model.
Source: I'm part of the training team
If you guys need help on GGUFs + Unsloth dynamic quants + finetuning support via Unsloth https://github.com/unslothai/unsloth on day 0 / 1, more than happy to help :)
Hey, really cool project, I’m excited to see the outcome.
Is there a blog / paper summarizing how you are doing it ?
Also which research group is currently working on it at eth ?
So you're not going to use copyrighted data for training? That's going to be a disadvantage with respect to LLaMa and other well-known models, it's an open secret that everyone is using everything they can get their hands on.
Good luck though, very needed project!
Not sure about the Swiss laws, but the EU AI Act and the 2019/790 digital millennium directive it piggies back on the topic, does allow for training on copyrighted data as long as any opt-out mechanisms (e.g. robots.txt) are respected. AFAICT this LLM was trained by respecting those mechanisms (and as linked elsewhere they didn't find any practical difference in performance - note that there is an exception to allow ignoring the opt-out mechanisms for research purposes, so they could make that comparison).
L3 has open pretraining data, it's just not official for obvious legal reasons: https://huggingface.co/datasets/HuggingFaceFW/fineweb
Are you using dbpedia?
no. the main source is fineweb2, but with additional filtering for compliance, toxicity removal, and quality filters such as fineweb2-hq
Thx for engaging here.
Can you comment on how the filtering impacted language coverage? E.g. finweb2 has 1800+ languages, but some with very little actual representation, while finweb2-hq has just 20 but each with a subdsantial data set.
(I'm personaly most interested in covering the 24 official EU languages)
we kept all 1800+ (script/language) pairs, not only the quality filtered ones. the question if a mix of quality filtered and not languages impacts the mixing is still an open question. preliminary research (Section 4.2.7 of https://arxiv.org/abs/2502.10361 ) indicates that quality filtering can mitigate the curse of multilinguality to some degree, so facilitate cross-lingual generalization, but it has to be seen how strong this effect is on larger scale
Imo, a lot of the magic is also dataset driven, specifically the SFT and other fine tuning / RLHF data they have. That's what has separated the models people actually use from the also-rans.
I agree with everything you say about getting the experience, the infrastructure is very important and is probably the most critical part of a sovereign LLM supply chain. I would hope there will also be enough focus on the data, early on, that the model will be useful.
When I read "from scratch", I assume they are doing pre-training, not just finetuning, do you have a different take? Do you mean it's normal Llama architecture they're using?
I'm curious about the benchmarks!
The infra does become pretty complex to get a SOTA LLM trained. People assume it's as simple as loading up the architecture and a dataset + using something like Ray. There's a lot that goes into designing the dataset, the eval pipelines, the training approach, maximizing the use of your hardware, dealing with cross-node latency, recovering from errors, etc.
But it's good to have more and more players in this space.
I'd be more concerned about the size used being 70b (deepseek r1 has 671b) which makes catching up with SOTA kinda more difficult to begin with.
SOTA performance is relative to model size. If it performs better than other models in the 70B range (e.g. Llama 3.3) then it could be quite useful. Not everyone has the VRAM to run the full fat Deepseek R1.
Disclaimer: I’m Swiss and studied at ETH. We’ve got the brainpower, but not much large-scale training experience yet. And IMHO, a lot of the “magic” in LLMs is infrastructure-driven.