The NSMnum syntax
It breaks up an NSMnum (Narrow Small Model [numerating]): one that does a single job and answers with a kind, a span, an item, a number or codes, never with sentences, and leaves the decision to code; narrow is the size of its answer, not of the model. It has 12 element types and 87 elements, and a NSMnum described in it is one choice of elements from each type.
NSMnum stands for Narrow Small Model (numerating); num is for numerating, because all this kind of model does is name a number: of a passage, of an item on a list written in advance, or none. I changed its short form from NSM to NSMnum on 15 November 2026 to avoid a collision: NSM is already the Natural Semantic Metalanguage, Anna Wierzbicka and Cliff Goddard's theory of meaning in linguistics, and the other short forms I tried were taken too. The same day the full name, until then Narrow Small Model, became Narrow Small Model (numerating), so that each part of NSMnum stands for a word. NSMnum is both singular and plural, written and said the same: one NSMnum, three NSMnum, layers of NSMnum, the nearest NSMnum already built; never NSMnums. Said aloud: en-ess-em-num (एन-एस-एम-नम), four syllables, the letters N, S and M and then num as in number, for one or many.
Instances
- ungli: a builder: design a task as steps of NSMnum (Narrow Small Models [numerating]) and plain code, and read beside each model what the nearest models already built were measured to do and what their records say to check. Its name is from ungli rakhna, to put a finger on the exact spot, which is all a model here does
Element types and elements
does holds one essential
The one thing the model does. A larger task is split into steps first, each done by its own model or by code.
- sorts
- puts each thing it reads into one of a few kinds set in advance
- points
- finds the part of what it reads that answers a question, or says there is none
- matches
- says which item on a list written in advance the thing it reads goes with
- spots
- says which of a set of words or sounds recorded in advance occurred
- scores
- gives each option a number, and code takes the highest
- chooses
- picks codes from sets, one slot at a time
- counts
- says how many of one kind of thing there are in what it reads shown in Lee and Lee, Journal of Real-Time Image Processing 21, 2024
- notices the unusual
- says how far what it reads is from the normal it learnt, so code can flag the odd case shown in Hammad, Iskandaryan and Trilles, Internet of Things 23, 2023
reads holds one essential
What goes into the model each time it answers.
- a sentence
- one sentence at a time
- two sentences
- two sentences read together, such as a question and its answer
- a passage
- a few sentences up to a page, with a question about it
- a whole text
- a whole text, shortened to fit the model's window
- a second of sound
- a second or two of sound from a microphone
- numbers
- a few whole numbers that describe a situation
- codes
- numbers that stand for words, never the words themselves
- a window of sensor readings
- a few seconds of readings from a sensor, such as a wrist's accelerometer shown in Di Leo, Biagetti, Falaschetti and Crippa, Sensors 25(12), 2025
- an image
- one photograph, or one frame from a camera shown in Gulzar, Sustainability 15(3), 2023
- a short video clip
- a few seconds of moving picture shown in Kondratyuk, Yuan, Li and others, CVPR 2021
answers with holds one essential
What comes out. Every answer can also be none, which code reads as don’t know.
- a kind
- one of a few kinds set in advance
- yes or no
- one of two answers
- a span
- where an answer starts and ends in what it read
- an item
- one item from a list written in advance
- a number
- a number for each option, which code compares
- codes
- codes chosen from sets, which code turns back into words
- a box
- where the thing is in an image, as four numbers, or nothing shown in Zagitov, Chebotareva, Toschev and Magid, Computer Optics 48(2), 2024
- several kinds
- every kind set in advance that is present, not just one shown in Su, Tomioka, Zhao and Liu, Image and Vision Computing 147, 2024
- a mask
- which parts of an image belong to which kind, pixel by pixel shown in Chae, Kim and Heo, LCTES 2024
- a stretch of time
- where in the sound or video something starts and ends shown in Lee, Kim, Song and others, IEEE Access 14, 2026
- points on a body
- where each joint of a person is, as numbers shown in Wang, Li, Cai, Chen and Han, CVPR 2022
starts from holds one essential
The trained model it begins from, before it learns the job.
- a pointing reader
- MiniLM, 33 million parameters, already trained to find the answer to a question in a passage
- an encoder with a new head
- an English encoder, such as MiniLM (33 million parameters) or DeBERTa, with a small new layer that picks one of a few kinds or says yes or no
- an adapter on a trained model
- a small set of added weights trained on top of a model already trained for a nearby job, a model that does not write text
- a multilingual encoder
- a small encoder that reads several languages, retrained on Indian languages
- a frozen audio encoder
- a model trained on sound, kept as it is, with a small scorer trained on top
- a keyword spotter
- a tiny encoder, tens to hundreds of thousands of parameters, that compares sound with recorded examples
- nothing
- no trained model: every weight is learnt from the start
- a few whole numbers
- no neural network: a handful of whole numbers that multiply the inputs
- a frozen image encoder
- a model already trained on many images, kept as it is, with a small new head trained on top shown in Gulzar, Sustainability 15(3), 2023
learns from holds one essential
Where the right answers it learns from come from.
- your own keys
- your own answers, given one item at a time on sheets that show no model’s answer
- a rule you stated
- answers code works out from a rule you stated in words
- rows a model wrote
- rows another model wrote for training, the right answer known from how each was made
- another model's labels
- labels another model gave, used on your word
- a dataset's labels
- the labels a public dataset came with
- comparisons you made
- which of two things you preferred, pair by pair, with no score asked of you shown in Hamilton, Tawn and Firth, Statistical Science 41(4), 2026
- the cases it asks you to label
- it picks the cases it is least sure of, and a person labels only those shown in Schröder, Müller, Niekler and Potthast, EACL 2023 demonstrations
- normal cases only
- examples of the ordinary, with no labels; what does not fit is the unusual shown in Hammad, Iskandaryan and Trilles, Internet of Things 23, 2023
- what it meets after it is set up
- it goes on learning on the device from new data, after it has been installed shown in Lin, Zhu, Chen and others, NeurIPS 2022
runs on holds one essential
Where the model answers once it is trained.
- a Raspberry Pi 5
- inside a toy or a game, on its processor, with nothing sent anywhere
- a phone's browser
- in the page itself, with nothing sent to a server
- a small server
- a server's processor with no graphics card, such as one with 1 CPU and 2 GB
- a graphics card
- a machine with a graphics card, such as an RTX 3060 with 12 GB
- a microcontroller
- a chip such as a Pico or an STM32, with a few hundred kilobytes of memory and no operating system shown in Di Leo, Biagetti, Falaschetti and Crippa, Sensors 25(12), 2025
data comes from holds one optional
Where the examples it trains on come from. Who gives the right answers is learns from. Optional: says where the examples it trains on come from, as in the NSMnum already built
- rows code made from templates
- rows that code builds from templates and word lists, with the answer known from how each was made
- texts a model wrote
- examples a language model wrote for training, read before use; the NSMnum itself never writes
- texts people wrote
- sentences or passages people wrote for some other purpose
- your own writing
- texts you wrote yourself, such as poems
- recordings made for it
- sound or sensor recordings made for this model, in the place it will run
- runs simulated by code
- games, mazes or other runs that code plays out to make examples
- pairs from a list written in advance
- pairs of items from a fixed list, such as a palette of blocks
- a public dataset
- examples someone else published, used only once you have agreed to the source. No NSMnum here has trained on one directly: rai's teacher reader read SQuAD 2.0 (126,689 rows) before teaching rai
how many holds one optional
How many examples it trains on. Optional: says how many examples it needs, so the keying can be planned
- a few dozen
- under a hundred, such as fifty poems keyed one by one
- a few hundred
- a hundred to a thousand
- a few thousand
- a thousand to ten thousand
- tens of thousands
- ten thousand or more, usually made rather than keyed one by one
tested on holds one optional
What it is tested on, kept out of everything it trains on. Optional: says what it is measured on and never trains on
- a held-out set from the same source
- examples like the training ones, never trained on
- your keys, given blind
- examples you marked on sheets before any model saw them
- sets keyed by hand
- sets marked by hand after they were written
- simulated runs
- runs that code plays out, compared with what plain code would do
- a published figure
- a result someone else published for the same kind of model; not a measurement of yours
- what people give it in use
- what people type, say or show it once it is set up. No NSMnum here has been measured on it yet; rai's record says it was not measured on answers typed by people who use it
code takes holds one optional
What the code works on. A step of plain code beside the NSMnum; each element names the NSMnum in ungli where code already does it. Optional: says what code works on, as in the NSMnum already built
- one NSMnum's answer
- the number, item or span one NSMnum gave, or its none
- several NSMnum' answers
- what more than one NSMnum gave for the same case
- the input itself
- the text, sound or numbers before any NSMnum reads them
- a list written in advance
- words, rules or a table people wrote before any case came
code does holds one optional
The one thing the code does with it. A step of plain code beside the NSMnum; each element names the NSMnum in ungli where code already does it. Optional: says the one thing code does with the NSMnum's answer
- checks
- keeps an answer only if it is valid, and counts anything else as none
- looks up
- finds the answer in a list written in advance
- counts
- tallies answers
- compares with a line
- turns a number into yes or no at a line people chose
- routes
- sends a case to the next layer, or to a person
- filters
- lets through only what passes a test
- assembles
- makes what a person sees from pieces written in advance
code gives holds one optional
What comes out of the code. A step of plain code beside the NSMnum; each element names the NSMnum in ungli where code already does it. Optional: says what comes out of the code, which is what a person meets
- a decision
- one of a few outcomes set in advance
- a name, or none
- who or what, from the text itself, or nobody
- a flag for a person
- a mark that a person should look, never a lock
- something made from pieces written in advance
- a picture, a move or a line of text assembled by code
- the input, passed on
- the input, or part of it, handed to the next step
- don't know, to a person
- the case no step could settle, handed to a person
Left out
A model that writes. Anything that writes text is not an NSMnum, and an NSMnum never starts from such a model or is taught by one: an NSMnum answers with a kind, a span, an item, a number or codes, and code decides what follows. Also left out: what the NSMnum is for in the world, its name, and the keys, which stay with the person who keyed them.
Described in it
- rai
- RAM
- Coherence v2
- aadha
- Hoja's gaali scorer
- Hoja Paarakhi's word spotter
- Beam's thread model
- Beam's maze maker
- aankh-micholi, the poem chooser
- Decay of Meaning word-sign model
- khatka
Small tasks
Tasks from ordinary work, each set out as a few steps of code and at most one NSMnum. None of them is built. Each opens in ungli, which shows beside it the NSMnum I have built for the most similar jobs, with what each was measured to do. An NSMnum that points hands back a place in the text, never words of its own.
A TA bill's journey dates
A travel allowance bill in a government office is paid for the days the tour was sanctioned for; any other day needs a sanction of its own. The dates the claimant typed are checked against the sanction order.
- An NSMnum: Points at each journey date in the claimant's typed journey details, or at nothing. Then code: Reads each date it was pointed at and compares it with the first and last day of the sanction order. A date outside them, or a bill where it pointed at nothing, goes to the accounts clerk.
Blocks: points · a passage · a span · a pointing reader · your own keys · a small server · one NSMnum's answer · compares with a line · a flag for a person
Purchase invoices against GSTR-2B
An accounts clerk claims input tax credit only on purchase invoices the suppliers have reported in their GSTR-1, so that they appear in GSTR-2B. Each invoice in the purchase register is matched with the GSTR-2B statement downloaded from the GST portal. No step needs a model.
- Code: Matches each invoice in the purchase register with GSTR-2B by the supplier's GSTIN, the invoice number and the tax amount, and lists the ones that do not match for the clerk.
Blocks: the input itself · looks up · a flag for a person
An RTI application sent to the right office
Under Section 6(3) of the RTI Act, an application asking for information another public authority holds is transferred to that authority within five days. The application is read for which authority holds what it asks for.
- An NSMnum: Chooses, from the office's list of public authorities, the one that holds what the application asks for, or none. Then code: If it chose this office, the application goes into the register of the office's Public Information Officer (PIO). If it chose another authority, code fills in the transfer letter written in advance. If it chose none, the PIO reads it.
Blocks: chooses · a passage · an item · a multilingual encoder · your own keys · a small server · one NSMnum's answer · routes · something made from pieces written in advance
A leave application against the attendance minimum
A student writes to ask for leave. Where the student's attendance would stay at or above the university's minimum, often 75 per cent of classes, the leave is recorded without anyone reading it.
- An NSMnum: Points at the first and last day of the leave asked for, or at nothing. Then code: Counts the classes on those days from the timetable. If the student's attendance stays at or above the minimum, the leave is recorded; if not, it goes to the faculty member. If it points at nothing, the application goes back to the student.
Blocks: points · a passage · a span · a pointing reader · your own keys · a small server · one NSMnum's answer · compares with a line · a decision
The syntax as data: nsmnum.json. It is read from ungli each time this site is built, so the two never differ.