Osye

AI & ML interests

None yet

Recent Activity

updated a model 1 day ago

Osye/co1

published a model 1 day ago

Osye/co1

liked a model about 1 month ago

s3nh/Bonsai-27B-unpacked-abliterated-uncensored

View all activity

Organizations

updated a model 1 day ago

Osye/co1

Updated 1 day ago

published a model 1 day ago

Osye/co1

Updated 1 day ago

liked a model about 1 month ago

s3nh/Bonsai-27B-unpacked-abliterated-uncensored

27B • Updated Jul 14 • 94 • 11

reacted to s3nh's post with 👍 about 1 month ago

Post

422

Uncensoring Mistral,

give it a try
s3nh/Ministral-3-14B-Instruct-2512-BF16-abliterated

liked a model about 1 month ago

s3nh/LFM2.5-350M-abliterated

Text Generation • 0.4B • Updated Jul 10 • 95 • 2

updated a dataset 2 months ago

Osye/gsm8k_pl

Viewer • Updated Jun 19 • 6.86k • 20

published a dataset 2 months ago

Osye/gsm8k_pl

Viewer • Updated Jun 19 • 6.86k • 20

reacted to s3nh's post with ❤️ 2 months ago

Post

468

Existing methods — GPTQ, AWQ, llama.cpp's k-quants — minimize empirical loss heuristically. None of them prove they are optimal in any information-theoretic sense. ICRB-Q builds a quantization scheme that is provably optimal via the Cramér-Rao lower bound (CRB): no unbiased estimator of a weight can have lower variance than [F(θ)]⁻¹, where F is the Fisher information matrix.