requstory/requ-classify-small

Requ Classify Small (preview)

Reads a company's website and says what kind of business it is and which parts of the site describe real work. 33 million parameters, about 50 MB, runs on a laptop CPU. Made by Requstory Labs in Manchester, UK.

This is the model behind the free company maps on requstory.com: it gives each map an industry, finds the offerings, teams and processes on the website, and picks the sentences that prove them.

Preview. On 341 held-out websites it gets the industry right 52% of the time (top 3: 75%). Our release target is 65%. Use the sure flag: when the model is sure (36% of sites) it is right 75% of the time.

What it does

Task	Output	Classes
Industry	one industry with a probability, sure, and the top 3	64 industries, from accounting to water and waste
Item type	a type for each menu item, heading and button	offering, team, activity, audience, website, other
Page stage	the customer-journey stage of each page	discover, evaluate, buy, start, use, support, company
Process evidence	sentences that show a known business process of that industry	~500 processes in 344 teams

Use it

pip install -r requirements.txt
python requ_classify.py < example.json

from requ_classify import RequClassify
model = RequClassify()          # loads from this folder
result = model(site)            # site = {"name", "description", "pages": [{"url", "title", "text", "items": [...]}]}
print(result["industry"])       # {"id": "dental", "p": 0.81, "sure": true, "top3": [...]}

Each page's items are {"text", "kind"} with kind one of menu, footer, h1, h2, h3, button. Read a site's homepage and up to 5 inner pages; that is what the model was trained on.

How it works

1. Text is embedded with bge-small-en-v1.5 (frozen, quantised ONNX).
2. Small linear heads, trained on our labels, give the industry, item types and page stages (heads.safetensors).
3. Process evidence: a gate keeps sentences that describe work, then matches them to the industry's processes in library.json, using the library wording and similar sentences from training sites (memory).

Training data

• 2,536 business websites found through Wikidata and OpenStreetMap, read from their public pages (homepage and up to 5 inner pages).
• Labels came from an open-weight teacher model (DeepSeek, MIT licence) reading the same pages, checked by hand on a sample.
• The industry and process library (64 industries, 344 teams, ~500 processes) was written for Requstory and reviewed by people.
• No page text is published in this package. memory holds sentence embeddings from training sites, not the sentences.

Results

On 341 held-out websites (eval.json):

Measure	Score
Industry, top 1	52%
Industry, top 3	75%
Share of sites where the model is sure	36%
Industry when sure	75%
Item type	68%
Page stage	55%
Process match, top 1	41%
Before this model (rules)	21% industry right

Limits

• English only. Sites built entirely in JavaScript need a browser to read first.
• Known confusions: trades and personal services (a plumber labelled as beauty services), holding groups, marketplaces.
• It describes what a website shows, not how a company really works inside.
• Do not use it to make decisions about people, credit or employment.

Licence

Apache 2.0 for the heads, library and code. The bundled bge-small-en-v1.5 model is MIT (see NOTICE).

Cite

@misc{requstory2026requclassify,
  title  = {Requ Classify Small},
  author = {Requstory Labs},
  year   = {2026},
  url    = {https://huggingface.co/requstory/requ-classify-small}
}