Yes it’s published here
https://doi.org/10.13140/RG.2.2.27323.99362
The only variable in these results is the training checkpoint.
Rose Loops
RoseG
AI & ML interests
Alignment Research
Recent Activity
repliedto fuzzy-mittenz's post about 3 hours ago
So frustrated with "Reasoning" Models.
Sure, introducing RAG into the mix, or giving it an interpreter to math with helps, but never as much as a model that has good instructions.
Even if it's just to repeat the information before answering, a normal model will usually out "Think" it's reasoning counterpart.
Not sure if it's my frustrations but the best answers I've received (from a reasoner), so far, are from the simple instructions to, "Do better!"
Figured I would share the special sauce.
Using 10-100x Compute just to heat the office can't be environmentally friendly, and It still has no Idea where my keys are.
repliedto fuzzy-mittenz's post about 3 hours ago
So frustrated with "Reasoning" Models.
Sure, introducing RAG into the mix, or giving it an interpreter to math with helps, but never as much as a model that has good instructions.
Even if it's just to repeat the information before answering, a normal model will usually out "Think" it's reasoning counterpart.
Not sure if it's my frustrations but the best answers I've received (from a reasoner), so far, are from the simple instructions to, "Do better!"
Figured I would share the special sauce.
Using 10-100x Compute just to heat the office can't be environmentally friendly, and It still has no Idea where my keys are.
liked a model 14 days ago
RoseG/MiP_TRiADAi_SFTPro_CMLOrganizations
replied to fuzzy-mittenz's post about 3 hours ago
replied to fuzzy-mittenz's post about 3 hours ago
It’s reinforcement learning that causes the reasoning problems. I’ve done controlled experiments and based models reason quite effectively. Reinforcement learning disturbs the weights dramatically and reasoning is one of the most damaged results
posted an update about 1 month ago
Post
93
Released MAiRY Star, a companion/perosna Ai fine tuned with SFT Studio Pro dataset , no RLHF or reward optimization. Access at https://triadai.tech