# MODA_NER

> An open benchmark and model family for fashion attribute extraction, by Hopit AI.
> Four tracks, never averaged. Predictions hashed before labels open. Losses published.

## What it is

MODA_NER is the MODA General Attribute Suite: four frozen evaluation tracks for extracting
structured attributes from fashion images and product text, plus the models measured on them.

Tracks, each with its own test set, metric and resampling unit:
- crop: 4,688 garment crops from 1,158 images, 15 fields, resampled by source image
- catalog: 9,995 catalogue images, 61,384 cells, 10 fields including colour and fit
- fullbody: 5,000 images across 1,751 product groups, 18 fields with an explicit not-applicable class
- text: 1,071 rows, 13 entity types in product copy

Scores are never averaged across tracks. Different images, fields and metrics.

## Results

crop attribute micro-F1 0.6300 against 0.6245, 95% interval [+0.0014, +0.0097]
catalog field-macro set F1 0.8292 against 0.6657, [+0.1595, +0.1676]
fullbody tier-1 macro-F1 0.6917 against 0.5943, [+0.0891, +0.1053]
fullbody not-applicable F1 0.6637 against 0.6088, [+0.0433, +0.0657]
fullbody visible macro-F1 0.5785 against 0.4969, [+0.0723, +0.0905]

The claim these support, in full: MODA General is the best of the named open systems evaluated
under the frozen public MODA General Attribute Suite, spanning localized garment crops, catalog
product images, color and fit, and applicability-aware full-body attributes. This is not a claim
of world-best, universal state of the art, human-gold quality, or production readiness.

On the crop track the comparator is our own architecture on a frozen third-party encoder, so
that row is an encoder ablation rather than a win over another vendor's product.

## Per-field reality

On crop, six of fifteen fields exceed F1 0.75 and can be populated automatically:
master_category 0.9215, category 0.8825, pattern 0.8356, sleeve_length 0.8073,
closure_type 0.7545, collar_presence 0.7508.

Nine need a human review step: sleeve_shape 0.6787, hemline 0.6428, sub_category 0.5966,
silhouette 0.5535, waist_type 0.5002, neckline 0.4650, collar_style 0.4566,
surface_treatment 0.4398, material 0.4148.

On 49.66% of garments all fifteen fields are correct.

## Models

moda-ner-v-crop, cropped garment input, MIT, commercial use permitted
moda-ner-v-catalog, catalogue image input, CC BY-NC 4.0, non-commercial only
moda-ner-v-fullbody, full-body photo input, CC BY-NC 4.0, non-commercial only
MODA_NER(T), product text, not distributed
MODA_NER Pro, hosted, contact https://hopit.ai

## Limits

All labels are dataset-native rather than independent human gold. A benchmark cannot prove its
taxonomy matches yours: on one external set a neckline vocabulary mismatch cost a production
route fourteen points with no model regression. And a strong extraction score does not transfer
to adjacent tasks — on mapping titles to a retailer's own product taxonomy, a terminal n-gram
baseline beats our text model.

## Links

Repository: https://github.com/hopit-ai/Moda_ner
Models: https://huggingface.co/HopitAI
Sibling project (fashion retrieval): https://github.com/hopit-ai/Moda