Source code, training pipeline, dataset and tests: https://github.com/alexsukhrin/uom-classifier (tag v0.4.0). This Hub repository contains the trained artifact and the dependency-free inference code only.

uom-classifier

Tiny, dependency-free classifier that maps noisy OCR'd Ukrainian unit-of-measure strings to canonical units — since v0.4 it also reads the product name of the same specification row.

from uom_classifier import UomClassifier

clf = UomClassifier()

clf.classify("пакува rhh")                      # ('упаковка', 0.9…)  ← OCR garble, model
clf.classify("пф", "Цефотрин пор. д/ін. 1г фл.")  # ('флакон', 0.8…)   ← name context helps
clf.classify("шт.")                             # ('штука', 1.0)      ← exact vocabulary hit
clf.classify("lilt.")                           # ('штука', 1.0)      ← exact reviewed form
clf.classify("шт/фл")                           # ('флакон', 1.0)     ← «штука» + one specific unit
clf.classify("<b>шт</b>")                       # ('штука', 1.0)      ← HTML stripped first
clf.classify("шт/уп")                           # None                ← unit + package dual, excluded
clf.classify("тис. доз")                        # None                ← carries an amount
clf.classify("порошок")                         # None                ← dosage form, not a unit
clf.classify("garbage")                         # None                ← __keep__ class / below threshold

None always means "leave the value as it is". The product name is optional (classify(raw) still works) and never overrides an exact answer.

How a string is resolved

Clean — HTML tags from the PDF text layer are stripped, whitespace collapsed.
Exact lookup — the vocabulary of common spellings, then the reviewed labeled forms (exact: 599 keys in the artifact, built from 854 reviewed forms minus vocabulary overlaps, amount-carrying forms and duals). Slash/paren forms whose halves name two different units («пач (штука)») never enter the table.
Pair rule — «штука» plus exactly one specific unit-class form («шт/фл», «шт. (амп.)») → that unit; two halves of the same canon («уп/упаковка») → that canon. Package pairs («шт/уп», «штука/контейнер») stay duals: they carry the price basis downstream.
Exclusions — the model must stay silent: slash and spaced duals, forms that carry an amount (a token starting with a digit or a multiplier — digits inside a letter token such as «д03», «уг1» are glyph confusion and are judged), one-letter strings, placeholders, dosage forms, service markers, units outside the model's classes.
Model — softmax regression over unit-string char n-grams (1..4) plus product-name features, with an explicit __keep__ class. It answers only when its best class is a unit and the confidence clears the threshold (0.78).

Product-name features: words as w:<first 5 letters> and ~30 curated cues (q:tablet, q:vol_l, q:count_no, q:reagent, …). The context vector is L2-normalized on its own and scaled by 0.5, so the unit string keeps the dominant voice. Training uses context dropout (35% of row copies without the name), so a missing name degrades to the form-only model.

Data (v0.4)

Real rows of Ukrainian public-procurement specifications (Prozorro), mined read-only from a production data-quality copy:

gold set (held out) — 464 rows / 246 forms, stratified by production row volume (all forms sampled per volume tier: ≥20 rows 50%, 10–19 50%, 4–9 40%, 2–3 35%, 1 row 20%; ≤6/4/3/2/1 rows per form), held out by form key from every training source (pairs, exact table, context rows). Labels: two independent judges (gpt-5.5, gpt-4o) saw the raw form, the product name and up to 4 sibling names — never an existing label; agreement on 255/464 (55%); the 209 disagreements went to a third judge (o3), who sided with gpt-5.5 in 132, gpt-4o in 34 and neither in 43. Headline rows: 421 (255 unanimous + 166 majority) — 243 canonical, 178 KEEP (unreadable 79, dual 57, not-a-unit 22, foreign unit 17, amount 15); 43 contested rows (no majority, 7.3% of weight) are excluded and reported. The file is frozen (data/gold.jsonl, sha256 627240f6…) before any training.
training rows — 1,406 deduplicated (form, product name) rows (weight 1,432): 436 rows labeled by the consensus of gpt-5.5 and o3, 477 more rows of forms whose consensus rows were unanimous, 406 rows with high-confidence campaign labels, 113 consensus rows of OCR forms from the extraction stage that never reached the item table. Judge agreement on training rows 74.3% (items) and 83.1% (extraction forms); disagreements are not trained on. 481 rows are KEEP.
755 form-only pairs and 125 vocabulary spellings (as in v0.3), 4,617 dictionary-labeled context rows (clean forms with product names).
LLM budget: 2,747 calls in total (gold 1,143; training 1,604).

Metric

A gold row is correct when the answer is its canonical label, or silence when the label is KEEP; wrong when a canon is answered that is not the label (any canon on a KEEP row); silence on a canonical label is a miss (neither). accuracy = correct / rows, weighted by the production row volume each gold row stands for; also reported form-weighted.

Results

Evaluated through the downstream service's resolution path (dictionary → this classifier with the product name).

	Gold accuracy	Gold wrong	Form-weighted acc / wrong
v0.3.0 as shipped (its exact table contains many gold forms — partly in-sample)	79.6%	9.85%	76.2% / 12.0%
v0.3.0 recipe, gold forms held out	66.3%	4.04%	62.3% / 4.8%
v0.4.0, gold forms held out	72.7%	2.88%	65.7% / 4.25%
v0.4.0 shipped (trained on gold labels too — in-sample, not a metric)	94.9%	1.16%	—

Held-out path breakdown (share of gold weight): pair rule 3.3% correct, model 17.2% correct / 2.9% wrong, silence 51.5% correct / 24.4% miss, dictionary after HTML stripping 0.7%.

Context ablation (same data v0.4-rc, gold held out): context 0.5 → 71.2% / 4.17%, no context → 70.1% / 4.71%; grouped CV +4–6 points on every seed.

Multi-seed CV (5 folds grouped by form key, gold held out, the whole resolution path simulated per fold, threshold tuned on the out-of-fold predictions of the first seed with a wrong-rate cap of 1.5%):

seed	accuracy	wrong
24948	65.9%	1.47%
1	63.6%	1.96%
2	63.6%	1.82%

The 95% / ≤2% target on unseen forms is not met. What blocks it (held-out gold, share of weight):

class	share	of which majority-only labels
model silent on an unseen form (`__keep__` / below threshold)	18.0%	50%
canon label on a unit+package dual («штука/контейнер» → контейнер) — excluded by policy	3.0%	60%
model wrong (mostly «іт» → штука vs `KEEP`)	2.9%	47%
canon label on a form carrying an amount («мг (по 250 мг)»)	1.6%	88%
other exclusions (one letter «ш», dosage form «спрей»)	1.0%	43%
dictionary-only canon («бан» → банка; no model class)	0.8%	0%

Even a perfect model under the current exclusions would reach 93.6% on held-out forms. The two judges themselves agree on 55% of gold rows.

On production traffic most forms recur: with the reviewed forms in the exact table, the shipped artifact resolves 54.7% of a week of live non-canonical rows (v0.3: 41.4%) and loses none of the gold rows v0.3 answered correctly.

Install

pip install git+https://github.com/alexsukhrin/uom-classifier

Retrain

pip install "uom-classifier[train]"   # numpy
PYTHONPATH=src python -m uom_classifier.train \
    --dataset data/uom_training_pairs.json --gold data/gold.jsonl \
    --out src/uom_classifier/data/uom_classifier.json

The trainer writes the shipped artifact and <out>_heldout.json (gold forms excluded everywhere) — evaluate the latter for honest numbers.

Dataset format (data/uom_training_pairs.json): canons, vocabulary, pairs (form → label, no context), exact (reviewed forms), rows (form, product name, label, weight = production rows, source), context_rows (dictionary-labeled clean forms with names). KEEP and units outside the model's classes train the __keep__ class.

Changelog

0.4.0

Product-name context features (artifact v3, context_scale), context dropout in training; classify(raw, product_name=None) stays backward compatible.
Explicit __keep__ class trained on judge-labeled garbage, duals and units outside the canon — the model learns to stay silent, not only the threshold.
New data: 1,406 labeled production rows with product names (two-judge consensus), frozen held-out gold set (464 rows), exact table 831 → 854 (48 campaign labels contradicted by the consensus dropped, 71 added).
Pair rule («шт/фл» → флакон, «уп/упаковка» → упаковка); unit+package duals never enter the exact table.
Digits inside a letter token no longer exclude a form — only forms that carry an amount do.
Sparse full-batch Adam training; form-grouped multi-seed CV that simulates the whole resolution path; threshold tuned on validation only.

0.3.0 — 24 forms promoted by an independent second judge; blister/jar/ canister stay out of the model.

0.2.0 — HTML stripping, exact table of reviewed forms (artifact v2), service exclusions.

0.1.0 — initial release.

Canonical units (22 model classes)

МО, ампула, г, доза, капсула, картридж, кг, контейнер, л, мг, мл, пакет, пара, пляшка, саше, таблетка, туба, упаковка, флакон, шприц, шприц-ручка, штука — plus __keep__.

License

MIT

Downloads last month: -; Downloads are not tracked for this model. How to track