aadha

aadha gives each sentence one of four marks, and writes nothing: 2 story, 1 the good noise, 3 common coin, 0 unwanted noise

release weights on Hugging Face NSMnum: Narrow Small Model (numerating) answers with one of four marks writes nothing base: MiniLM-L12-H384 open in ungli code: AGPL-3.0 weights: CC BY 4.0

A small reader that gives each sentence one mark:

mark means
0 unwanted noise
1 the good noise
2 story
3 common coin

The good noise is writing with more than the usual amount of abstraction, layered meaning or analogy, where it says something a reader could argue with. Common coin is a pairing already in circulation, a saying or a stock phrase. Story is plain meaning. Unwanted noise is writing that holds no ground: decoration, nonsense, or a figure that fits anything.

aadha only marks. It writes nothing. Code, not the model, decides what the marks mean for a piece of writing. In teetar aadhi bater, the tool built on it, the page shows how the marks are spread across the piece. Its name is from the Hindi idiom aadha teetar aadha bater, half partridge, half quail: a thing that is neither one thing nor the other.

It is an NSMnum (Narrow Small Model [numerating], said en-ess-em-num, एन-एस-एम-नम; written NSM until 15 November 2026, when the short form was changed to avoid a collision with the Natural Semantic Metalanguage in linguistics). What is narrow is its answer, which is always one of four marks; the model's own size does not decide it. It is an exemplar in ungli, the builder for NSMnum, and part of SYNTAX.

One sentence goes to aadha, which answers with one mark, 0 to 3, and nothing else; code decides what the marks mean

Use it

The weights are on Hugging Face at prayasabhinav/aadha.

pip install torch transformers peft
python use.py "The bus was late again." "Every mistake leaves a fingerprint on the next one."
python use.py --reflect "..."

--reflect loads the reflect adapter, a second reader trained on top of the first for argued or reflective pieces. Nothing chooses between the two for you: the person using the tool chooses.

How it reads a page

Every sentence of my Museum of Vestigial Desire page Music (2023), one sentence per cell, as the released model marked them on 15 November 2026:

80 sentences of Music in two rows: the plain reader marks 78 story, 1 common coin and 1 unwanted noise; the second reader marks 57 story, 20 the good noise and 3 unwanted noise

Two sentences from the page, as the plain reader and the second reader marked them:

Narratives have musical arcs. Music has narrative arcs. Each is story to the plain reader and the good noise to the second reader

How it is made

microsoft/MiniLM-L12-H384-uncased with a new four-way head, trained three passes at 5e-5, batch 16, on three seeds whose weights are averaged into one ("souped"). Sentences are read up to 96 tokens. The rows are made by the scripts in scripts/ (gen.py and the row builders beside it), and train.py trains an arm. The reflect adapter is a LoRA adapter trained with adapter.py.

The version released here is the one last served, o4x3 with the reflect adapter, live from 3 November 2026 until the reader was taken offline on 9 November 2026. Its training labels are Claude's reading of the rows, which I agreed to use; they are not my keys. Every test it was measured on I keyed myself, before it was trained. The training rows themselves are not released.

What it was measured to do

result tested on
all sentences 576 of 603 marked as I marked them (Haiku 441, Sonnet 456) 603 sentences I marked before any model saw them, 30 October 2026; an earlier version, not the one released here
the good noise 62 of 72 found (Haiku 72, Sonnet 67) the same 603 sentences, the same earlier version
present marks 176 of 182 right sets I keyed under the present marks, 3 November 2026; the version released here

Hosted models matched or beat aadha on the good noise. Its whole lead over them was on one boundary, abstract nonsense, which they called the good noise. Two trainings of the same setup, each averaged over three seeds, disagree on 13 to 16 sentences of the large test sets, so a difference smaller than that means nothing. Two things are not solved: common coin in fresh wording (0 of 6, for every version), and misspelt story, which can read as unwanted noise.

Bars: on all 603 sentences an earlier aadha marked 576 as I did, Claude Sonnet 456, Claude Haiku 441; on the 72 good-noise sentences Claude Haiku found 72, Claude Sonnet 67, aadha 62

Licences

The code is under the GNU Affero General Public License 3.0 (LICENSE). This text and the model card are under CC BY 4.0. The weights are under CC BY 4.0. By Prayas Abhinav.

Downloads last month
64
Safetensors
Model size
33.4M params
Tensor type
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for prayasabhinav/aadha

Finetuned
(154)
this model