Repository navigation
aadha 0.8.1 is the version last served, o4x3 with the reflect adapter. The code is here and the weights are on Hugging Face at prayasabhinav/aadha.
It gives each sentence one mark: 0 unwanted noise, 1 the good noise, 2 story, 3 common coin. It writes nothing; code decides what the marks mean for a piece of writing.
On the sets I keyed under the present marks, it marked 176 of 182 sentences as I did (3 November 2026). Its training labels are Claude's reading of the rows, which I agreed to use; they are not my keys. The README gives every result and what each was measured on.
It is an NSMnum (Narrow Small Model [numerating]), said en-ess-em-num (एन-एस-एम-नम): the letters N, S and M, then num as in number. The short form was NSM until 15 November 2026. You can design a task around it, or around any NSMnum, in ungli.