Weaver: New Specialised Writing LLMs Outperform GPT-4 (arxiv.org)
from hexual@lemmy.world to technology@lemmy.world on 06 Feb 2024 20:46

Weaver introduces a new family of specialised large language models tailored for creative and professional writing. Offering models ranging from 1.8B to 34B parameters, said to outperform larger generalist models like GPT-4 by focusing on human-like text production and diverse content creation capabilities.


threaded - newest

toothbrush@lemmy.blahaj.zone on 06 Feb 2024 21:58 next collapse

is this an open source AI?

abhibeckert@lemmy.world on 06 Feb 2024 22:22 next collapse

I don’t think they’ve said what the license will be.

FaceDeer@kbin.social on 06 Feb 2024 22:24 next collapse

One of the size classes they mention in the abstract is called "Weaver Pro" so my initial assumption would be that it's not. However, I find that with this sort of thing the most important secret is that something is possible. If Weaver works as advertised we will now know that it's possible fir a 34B model to get better-than-GPT4 performance, which means lots of people will be willing to devote resources to recreating it since they now know those resources won't be wasted.

And if Weaver is meant to be "commercial" I wouldn't be surprised if there's a bunch of censorship baked into it, so the eventual open-source version will have an advantage.

FunderPants@lemmy.ca on 06 Feb 2024 22:32 collapse

It doesn’t seem to be. Their Chinese website talks about buying AI credits, their English website only has a waitlist but this looks more like a new closed commercial product than anything else.

Also, check the appendix in the paper, I think it’s a bit concerning that the second author is responsible for the writebench benchmark they use to make their claims about the model. That is, the evaluation isn’t independent from the authors.

I mean, I’m not saying they’re not right, just that this is a yellow flag to investigate more.

Second flag is I don’t see a journal this will/is published in. Arxiv is not peer reviewed.

A. Appendix A.1. Author Contributions Tiannan Wang is the core contributor of Weaver. Tiannan is responsible for continual pre-training, supervised fine-tuning, and preference optimization. Tiannan is also a main contributor for the data synthesis and the benchmark/evaluation process.

Jiamin Chen is a main contributor of Weaver. Jiamin is responsible for WriteBench and is also main contributor for data synthesis and model evaluation process

BetaDoggo_@lemmy.world on 06 Feb 2024 22:10 collapse

Seems kind of like phi but for writing, the smaller ones are trained with 50B tokens and the largest is only trained with 18B.