Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Mistral's Shieldstral: 3B open-weights model for multimodal moderation

mistral.ai

Post reply on HN
Print view View original

51–60 of 154 posts

Re: Mistral's Shieldstral: 3B open-weights model for multimodal moderation

#51

Post by lucrbvi »

himata4113 wrote: post #28
fastball wrote: post #7

Should've called it Safestral. Also I do like Mistral's seemingly newer strategy of focusing on smaller, more fine-tuned models for various use-cases, presumably the result of their large MoE models not competing effectively with the frontier models.

It's not that their strategy is to train smaller models, it's the only choice they have. Training SOTA takes anywhere from 1.5b to 150b. We don't know the real cost of training for the chinese models, but mistral neither has the compute nor money to do that.

Mistral has the capability of training such models. Take a look at Poolside[1], they are claiming to pre-train their Laguna series of models on 4,096 NVIDIA H200 GPUs[2]. Mistral has approximately 13,800 NVIDIA GB300 GPUs, which are nearly 2x more efficient for training.

The problem with Mistral is that they do not seem to have aligned incentives to train big open-weight models, even if the teams would like to.

[1]: https://poolside.ai/ [2]: https://poolside.ai/blog/introducing-laguna-s-2-1

Re: Mistral's Shieldstral: 3B open-weights model for multimodal moderation

#53

Post by trilogic »

This model is way small for a proper assessment (imo). It should be very useful to study how big the real model must be for this purpose. Maybe merging it to a bigger one (adding it as expert style in moe) would be a solution! Great job to Mistral team.

Re: Mistral's Shieldstral: 3B open-weights model for multimodal moderation

#54

Post by gizmodo59 »

how does this compare with https://developers.openai.com/api/docs/models/omni-moderatio...

As for use cases, obviously we can't fully rely on non-deterministic capability for sensitive things but a small model which can do a good job acts as a first defense and then a human can review later.

Re: Mistral's Shieldstral: 3B open-weights model for multimodal moderation

#55

Post by mosura »

cyanydeez wrote: post #43
mosura wrote: post #8

Someone should use this to do the exact opposite of the intention: filter for “offensive” content, and boost it or collate it into a newsletter/email blast for people of culture. You have to give it to Mistral they do at least know what the market near them says they want right now. The great problem is in a few years of this that market won’t be worth anything. Edit to add, you could also add this to an AI workflow…

[flagged]

The bad words police are the actual authoritarian nutjobs. That is the problem.

Re: Mistral's Shieldstral: 3B open-weights model for multimodal moderation

#56

Post by grezql »

kergonath wrote: post #41
baq wrote: post #31

Earlier quoted context omitted.

Note I didn't mention copyright

By which other mechanism could American AI companies prevent this? Other companies don’t really care about EULAs and even if they needed to care it’s trivial to let third parties do it. Why would they? Almost nobody in the space cares about copyright and play fast and loose with laws and regulations. What’s the mechanism that could today prevent other companies from using LLM outputs to train their models?

[dead]

Re: Mistral's Shieldstral: 3B open-weights model for multimodal moderation

#57

Post by moffkalast »

fastball wrote: post #7

Should've called it Safestral. Also I do like Mistral's seemingly newer strategy of focusing on smaller, more fine-tuned models for various use-cases, presumably the result of their large MoE models not competing effectively with the frontier models.

What, you don't want a model called the Shitstral-3B :D

Re: Mistral's Shieldstral: 3B open-weights model for multimodal moderation

#59

Post by charcircuit »

hypfer wrote: post #15

I would be curious if this can do moderation with an arbitrary ruleset, or if it's just "that one moderation style" we already know from current big tech platforms. The kind where malicious intent is okay if the words are nice. ___ Or, rephrased: How big is the space in which you can tune this model without retraining. Is it just "we hate sex"/"we don't hate sex" "We hate violence"/"we don't hate violence" or is it _…

It sounds like it is. You have a set of moderation policies and then you evaluate the model 1 time per policy if it is violating it. Then you combine the results into a score you use for taking actions off of.
Post reply on HN