LinuxCommandLibrary
GitHubF-DroidGoogle Play Store

janas-nsm

Write and score Natural Semantic Metalanguage explications

TLDR

Explain one word with a model trained for it
$ janas-nsm [model.jns] [chair]
copy
Pass example sentences the model should see
$ janas-nsm [model.jns] [chair] ["a wooden chair by the window"]
copy
Explain a file of words and keep the JSON
$ janas-nsm [model.jns] --batch [words.jsonl] --out [out.jsonl]
copy
Score explications that already exist, with no model
$ janas-nsm --score [explications.jsonl]
copy
Use the paper's prompt and the full prime list
$ janas-nsm [model.jns] [leg] --prompt paper --lexicon full
copy
Let the model write only primes, function words, and molecules
$ janas-nsm [model.jns] [leg] --strict -n [5]
copy

SYNOPSIS

janas-nsm model.jns word [example...] [options]janas-nsm model.jns --batch file [options]janas-nsm --score file [options]

DESCRIPTION

janas-nsm asks a model trained for the Natural Semantic Metalanguage (NSM) for a word's explication, then counts what that explication is made of. NSM paraphrases a meaning with a small set of semantic primes (someone, something, do, happen, want, know, good, before, and the rest) plus a few simpler words. The model Janas was checked with is DeepNSM (Llama 3.2 1B with a LoRA adapter, [arXiv 2505.11764](https://arxiv.org/abs/2505.11764)), converted with hf2gguf and gguf2jns. Janas does not ship those weights.One word prints the explication, then a line of counts. A batch reads JSON lines and writes one JSON line per sample. --score only counts: each input line already has the explication, and no model is loaded.The default prompt is the model card's (`Word:`, `Examples:`, `Paraphrase:`). --prompt paper uses the indented prompt from the paper's evaluation script. Sampling follows DeepNSM: temperature 0.7, top-k 50, top-p 0.95, at most 512 tokens. Speculative decoding is off.Counts use one of two prime lists. paper (the default) is the list DeepNSM's published scores were computed with, single words, without "i" and "said". full is Goddard and Wierzbicka's 65 primes, their inflected forms, and their phrases. Other words are function words or molecules. Legality is 10 times (primes minus molecules) divided by the number of words, from -10 to 10. An explication is circular when one of its words has the target word's English lemma (suffix rules, not spaCy). On the paper's 1,341 published records, circularity matches spaCy on every record, and primes and molecules match on 1,337.A batch ends with means and standard errors on the standard error, plus read and write speed when a model ran. Progress prints every 10 words unless --quiet.

PARAMETERS

model.jns

Janas model file. Required unless --score.
word [example...]
The word to explain, then optional example sentences. Required for a single word. Ignored when --batch or --score is set.
--batch file
JSON lines. Each object has "word" and "examples" (an array of strings). At most 64 examples from a line are used.
--score file
JSON lines with "word" and "explication" (or "text"). Counts them and loads no model.
--out file
Where a batch's JSON lines go. Default is the standard output. The summary stays on the standard error.
-n count
Explications per word (default 1). Must be at least 1.
--seed n
Sampling seed (default 1).
--prompt card|paper
card (default) or paper. Anything else is an error.
--lexicon paper|full
Prime list. full selects the 65 primes with forms and phrases. Any other value, including paper, selects the paper's list.
--strict
The model may write only prime forms, function words, punctuation, and the molecules. The output head computes those tokens alone.
--molecules file
Molecules --strict allows, one word a line. Default is a built-in list (body parts, physical qualities, nature, people, everyday actions, and things). An empty file allows none.
--temp t
Temperature (default 0.7).
--top-k k
Top-k (default 50).
--top-p p
Top-p (default 0.95).
--max tokens
Most tokens in one explication (default 512). Must be at least 1.
--quiet
No progress on the standard error. A single word still prints its explication.
-h, --help
Print usage on the standard error and exit 0.

CAVEATS

DeepNSM's weights, code, and data are CC BY-NC-SA 4.0. Convert a checkpoint with hf2gguf for research; do not expect Janas to include the model.Context length is fixed at 4096 tokens. --strict restricts the output head on the CPU. A pass that runs entirely on the GPU still computes the whole head and applies the vocabulary afterwards. The restricted rows are not used on the GPU or the NPU.--strict can raise legality and speed and still make the explication worse at saying which word it is. On DeepNSM-1B Q8_0, five seeds each way over the paper's 149 benchmark words, legality moved from 5.33 to 5.48 while a second model ranked the target first less often (81.9% down to 75.8%). The default writes freely.A batch's JSON has "word", "sample", "explication", "legality", "primes", "molecules", "words", "circular", "tokens", and "seconds". --score writes "word" plus the counts only. Single-word mode prints the text, then a bracketed summary, not JSON.Exit 2 is usage (missing arguments, a bad prompt, or a count below 1). Exit 1 is a missing file, bad JSON, or a generation error. Unknown options are exit 2.

HISTORY

janas-nsm was added on 25 October 2026, in the same Janas release as Llama 3 and LoRA merge support in hf2gguf. It ships with Janas, written by Maurizio "camauri" Cammalleri (first public release September 2026, GNU GPL 3 or later).

SEE ALSO

RESOURCES

Copied to clipboard
Kai