DaisyRose-27B

140 points - yesterday at 7:39 PM

Source

Comments

woah yesterday at 9:12 PM
The heavy reliance on eval datasets showcased here and also evidenced by the frontier lab's high payments for such data cast doubt on the claims that AI will become self-improving soon
jacquesm yesterday at 9:28 PM
I always wonder what you lose when you start maxing out some specific benchmark. It is really hard to do this in a balanced way and I would love if everybody that did these kind of efforts would take the time to also list the before/afters of a number of other benchmarks that they did not use in their training.
pkoiralap yesterday at 10:40 PM
Sabi-noema also looks like a close competitor in the same space. Built from qwen3.8 and made better.

https://huggingface.co/sabicap/sabi-noema

Also HF model here: https://huggingface.co/RadixArk/DaisyRose-27B

xlayn yesterday at 8:34 PM
Snap, no gguf to download, will wait for one to appear to test, any improvement over an already good model is good news.

"18 days on 144 H200 GPUs" insane.. didn't thought so much hardware would be needed to train Qwen.

mips_avatar yesterday at 11:06 PM
How did you guys set up the RL topology? Were you guys doing individual RL nodes (4xh200 rollout/4xh200 train) or did you have dedicated trainer and rollout nodes?
deleted yesterday at 7:47 PM
0x69420 today at 1:20 AM
3.8 27B is an interesting choice. on the face of it, doing RL on it feels like refrying twice-cooked pork then stuffing it in the microwave for good measure. the fact that even one benchmark (tb4.1) increased more than marginally is kind of miraculous. would be interesting to see where "miles" would take eg. 3.8 flash next, which is already impressive despite being an under-trained architectural preview.
robertkarl yesterday at 8:28 PM
I look forward to testing this against my private eval set to determine if it's benchmaxxing or actual improvement based on just 1200 tasks. downloading it now
XCSme yesterday at 8:35 PM
Is it post-training or bench-maxxing?
octoberfranklin yesterday at 8:43 PM
FWIW that is about $331,000 on low-onboarding-grief clouds (e.g. vast.ai) right now using the SXM variants they used.

You might be able to shave that down using a different flavor of H200 (like PCIe) but it would take more hours, so benefits are unclear.

amarinhodev yesterday at 10:04 PM
[flagged]