spellman model

Runtime model for spellman — a Cyrillic-optimized language detector. 30 classes: the G10 world languages plus ~20 Cyrillic-script languages, built for the regime general-purpose detectors fail at — closely-related Cyrillic pairs (ru/be/uk, bg/mk/sr, kk/ky/tt/ba) on wild, short, real internet text — at single-digit µs/sample CPU speed.

Languages (30)

• Cyrillic group (21): rus ukr bel bul mkd srp kaz kir tgk uzn tat bak chv sah tyv mon oss che udm mhr kpv
• Latin group (5): eng spa fra por deu
• Script-routed, no model columns (4): cmn jpn hin ara

Files

• model.json — the runtime contract (version: 3): language order (model columns), hash id/seed, char n-gram config (1..=5), canonicalization + lexical flags, calibrated confidence threshold θ.
• model.safetensors — P [131073×30] f16 (the algebraic fold P = E·W of a fastText-style model) + bias [30] f16. Inference is pure table lookups — no embedding gathers, no matmul.

The feature space is char n-grams plus a lexical channel: every word also contributes its whole-word FNV-1a-64 key and an adjacent-word bigram key (fastText's word / wordNgram features) hashed into the same bucket space.

Storage formats

The same model in every folded-table storage format; the runtime dequantizes at load and the graph is unchanged — every variant is accuracy-identical to the f16 original within ±0.02pp.

ref	format	size
vpermilp/spellman (root)	f16	7.86 MB
vpermilp/spellman/int8-row	int8, per-row scales	4.46 MB
vpermilp/spellman/int8-col	int8, per-column scales	3.93 MB
vpermilp/spellman/fp8-row	fp8 e4m3, per-row scales	4.46 MB
vpermilp/spellman/fp8-col	fp8 e4m3, per-column scales	3.93 MB

The spellman CLI fetches any of them straight from the Hub:

spellman detect --model hf:vpermilp/spellman/int8-col < text.txt

Usage

This is not a standalone checkpoint: the featurizer (canonicalizing char n-grams + lexical word n-grams + signed feature hashing) lives in the runtime, so the model is consumed by the spellman Rust workspace (detector crate + CLI):

hf download vpermilp/spellman --local-dir model
echo "Съешь ещё этих мягких французских булок" | spellman detect   # rus
printf 'Қазақша\nHello\n' | spellman detect --lines               # kaz / eng

Metrics

eval	accuracy
held-out mix (85,283 rows, pristine test)	98.28%
Tatoeba (37,051, out-of-domain)	98.66%
Tatoeba granularity ladder (word / pair / triple)	68.5 / 87.1 / 94.1
wild Russian tweets (2,679; 61% @mentions, 20% URLs)	93.73%
throughput	~3.5 µs/sample (fp16 JIT plan, CPU)

On identical eval rows it beats fastText lid.176 (Tatoeba 94.90% → 98.66%) and lingua 1.8 high-accuracy on the shared Cyrillic close pairs at ~30× its speed; whichlang and lingua cover only 10 and 17 of these 30 classes respectively. Against open-LID SOTA GlotLID v3 it wins the wild held-out workload (98.28 vs 96.43) and the single-word rung (68.5 vs 43.9) while running ~100× faster; GlotLID's much larger training set still wins clean out-of-domain sentences (99.25 vs 98.66).

Training data

FineWeb-2 per-language configs sampled as line-windows (ODC-BY) + FineWeb (English), ~104k Tatoeba sentences (CC BY 2.0 FR), and per-language community/parallel corpora for the weak classes (Tatar, Bashkir, Chuvash, Tuvan, Kyrgyz, Udmurt, Mari, Macedonian, Chechen, …), hygiene-filtered with twin-language protection. Training methodology and the full recipe are documented in the spellman repository (design doc + training guide).