Didn’t see any posts about these models so I made one myself.
This first set of models was trained with 288B high quality tokens, will be interesting if the 51B and 102B models hold up. Commercial use is allowed with no authorization.
https://github.com/IEIT-Yuan/Yuan-2.0/blob/main/README-EN.md
(Chinese) https://github.com/IEIT-Yuan/Yuan-2.0
Paper: https://arxiv.org/abs/2311.15786
Huggingface download links
https://huggingface.co/pandada8/Unofficial-Yuan-2.0-2B
https://huggingface.co/pandada8/Unofficial-Yuan-2.0-51B
https://huggingface.co/pandada8/Unofficial-Yuan-2.0-102B
Here’s the second set of models I found. 7B and 65B were trained with 2.6T tokens, and the 13B with 3.2T. The 65B model supports up to 16K context, while the two smaller ones support up to 8K.
https://huggingface.co/xverse/XVERSE-65B
https://huggingface.co/xverse/XVERSE-13B
https://huggingface.co/xverse/XVERSE-7B
These models know 40 over human languages plus several programming languages too. Commercial use is allowed, but you have to submit an application form.
So I don’t know much about architecture but I’m assuming if we want to run something like this in Llama, we’re going to need to submit a request? If its ground up, then pretty much everything is going to need to be implemented, right?
- Deepseek 67B still beats XVERSE-65B in the benchmarking scores.
- The benchmarks indicate strong math and coding performance for these two model series.
- Yuan has a unique optional attention mechanism that enhances output quality
I’m really interested in having a 51B model. I would love something between 34B and 65/70B.