The doom is still buried within Command-A for sure.

The doom is still buried within Command-A for sure.

Only another 38 days to go:

A step 601 preview - all with temperature = 0:

https://pastebin.com/GASKaHTk

https://pastebin.com/CRT81QLb

• It's still messing up some end of lines, but I can live with that if it works... Likely can be fixed later using the new class 0 random data if a problem.
• The Grimdark story was noticeably (much!) better compared to the inverse.
• The Battlestar Galactica story showed that even though Q8_0, F16 and BF16 all diverge slightly from F32; it's not clearly making them any worse (I actually liked the Q8_0 story best!).

It still has a way to go before it starts to converge, but I would think by step 1000 it will be pretty close:

566 responses in previous thread! In the future we may be the reason for hf staff to implement multi-page view of discussions.

This was posted on Hacker News today:

https://outsidetext.substack.com/p/how-does-a-blind-model-see-the-earth?selection=5413dcae-b9f4-4adb-8826-d48e3908de2a#:~:text=Wow%2C%20best%20rendition%20of%20the%20Global%20West%20so%20far

Absolutely fascinating!

This was posted on Hacker News today:

https://outsidetext.substack.com/p/how-does-a-blind-model-see-the-earth?selection=5413dcae-b9f4-4adb-8826-d48e3908de2a#:~:text=Wow%2C%20best%20rendition%20of%20the%20Global%20West%20so%20far

Absolutely fascinating!

That was really cool. Thanks for sharing!

This was posted on Hacker News today:

https://outsidetext.substack.com/p/how-does-a-blind-model-see-the-earth?selection=5413dcae-b9f4-4adb-8826-d48e3908de2a#:~:text=Wow%2C%20best%20rendition%20of%20the%20Global%20West%20so%20far

Absolutely fascinating!

That was really cool. Thanks for sharing!

Yeah, and llama-3.1:405b doing so well was quite a surprise too (and makes you a bit sad everything seems to be moving away from large dense models ).

It is a real base, but it can't hold a conversation, just like its predecessor. I have honestly no hope for fully open models. Copyrighted data is best data, and lack thereof is noticeable.

but was very dumb, incapable of most basic stuff

I suspected as much. Nice that they've left all the earlier checkpoints in the repo history.

If you still have the pytorch weights and want to convert it, this fork worked for me:

git clone -b k2v2 https://github.com/cturan/llama.cpp.git llama_k2v2
cd llama_k2v2
pip install -r requirements/requirements-convert_hf_to_gguf.txt
python convert_hf_to_gguf.py --outtype q8_0 /workspace/K2-V2/

I didn't actually try running it though.

Yeah, I can't say I'm all that impressed with it either

It's started giving me the shits with the reasoning about safety policy.

What did you think of Devstral-123B ?

To me the vibes are Devstral-123B is to Largestral what Command-A is to Command-R+

For creative writing:

Largestral-2407 > Devstral > Largestral-2411

But keep in mind, I get the impression that I like Command-A more than you guys do lol

I tri d out Devstral 123b on NanoGPT and was surprised how good it was at RP and figured I must have just got lucky 🤣

Happy New Year!

Happy New Year!

Happy New Year!

Happy new year!

https://github.com/sionic-ai/solar-vs-glm
solar in shambles?

https://www.reddit.com/r/LocalLLaMA/comments/1q31ltd/local_llms_vs_breaking_news_when_extreme_reality/
This is why I absolutely hate stubborn LLMs.