4 ms·
I hope they do well. AFAIK they’re training or finetuning an older LLaMA model, so performance might lag behind SOTA. But what really matters is that ETH and EP
by isusmelj 1y ago
I hope they do well. AFAIK they’re training or finetuning an older LLaMA model, so performance might lag behind SOTA. But what really matters is that ETH and EPFL get hands-on experience training at scale. From what I’ve heard, the new AI cluster still has teething problems. A lot of people underestimate how tough it is to train models at this scale, especially on your own infra.
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.
- lllllm 1y agoNo, 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
- d3m0t3p 1y agoHey, 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 ?
- blurbleblurble 1y agoAre you using dbpedia?
- lllllm 1y agono. the main source is fineweb2, but with additional filtering for compliance, toxicity removal, and quality filters such as fineweb2-hq
- PeterStuer 1y agoThx 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)
- lllllm 1y agowe 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 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
- danielhanchen 1y agoIf you guys need help on GGUFs + Unsloth dynamic quants + finetuning support via Unsloth https://github.com/unslothai/unsloth https://github.com/unslothai/unsloth on day 0 / 1, more than happy to help :)
- lllllm 1y agoabsolutely! i've sent you a linkedin message last week. but here seems to work much better, thanks a lot!
- danielhanchen 1y agoOh sorry I might have missed it! I think you or your colleague emailed me (I think?) My email is daniel @ unsloth.ai if that helps :)
- isusmelj 1y agoThanks for clarifying! I wish you all the best luck!
- Al-Khwarizmi 1y agoSo 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!
- badsectoracula 1y agoNot 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).
- miraculixx 1y agoThat is not correct. The EU AI Act has no such provision, ans the data mining excemption does not apply as the EU has made clear. As for Switzerland copyrighted material cannot be used unless licensed.
- moffkalast 1y agoL3 has open pretraining data, it's just not official for obvious legal reasons: https://huggingface.co/datasets/HuggingFaceFW/fineweb https://huggingface.co/datasets/HuggingFaceFW/fineweb
- menaerus 1y agoWait, whole (english speaking) web content dataset size is ~50TB?
- zX41ZdbW 1y agoYes, if we take the filtered and deduplicated HTMLs of CommonCrawl. I've made a video on this topic recently: https://www.youtube.com/watch?v=8yH3rY1fZEA https://www.youtube.com/watch?v=8yH3rY1fZEA
- menaerus 1y agoFun presentation, thanks! 72min ingestion time for ~81TB of data is ~1TB/min or ~19GB/s. Distributed or single-node? Shards? I see 50 jobs are used for parallel ingestion, and I wonder how ~19GB/s was achieved since ingestion rates were far below that figure last time I played around with CH performance. Granted, that was some years ago.
- zX41ZdbW 1y agoDistributed across 20 replicas.
- asjir 1y agoI'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.
- zettabomb 1y agoSOTA 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.
- tough 1y agoalso isn't DeepSeek's Mixture of Experts? meaning not all params get ever activated on one forward pass? 70B feels like the best balance between usable locally and decent for regular use. maybe not SOTA, but a great first step.
- luke-stanley 1y agoWhen 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!
- andy99 1y agoImo, 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.
- alfalfasprout 1y agoThe 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.