ttucs-model-bow

These are the pretrained weights for ttucs, the Text-to-UCS slate classifier. The package contains data only: config.json, fp16 weights.safetensors and a SHA-256 manifest.json. It has no code dependencies and makes no network calls.

You normally don't install it directly: pip install ttucs pulls it in. ttucs then finds it offline:

from ttucs import TTUCS

model = TTUCS.from_pretrained()          # resolves the installed ttucs-model-bow weights
print(model.predict("wooden door creak slow, take 3").cat_id)

ttucs info --verify shows where the weights were loaded from and checks them against the manifest.

Model

• Architecture: hashed word 1–2-gram + char 3–5-gram bag of embeddings (131,072 buckets × 256 dims), a residual MLP, then UCS CatID and Category heads. Temperature-calibrated. 738 UCS v8.2.1 CatIDs.
• Training data: synthetic descriptions and slates generated from the UCS v8.2.1 list, plus UCS label documents. No third-party library data is used, so the weights can be redistributed.
• Metrics and limitations: see the model card.
• License: MIT (weights). The UCS list is a public-domain initiative by Tim Nielsen, Justin Drury, Kai Paquin and contributors (https://universalcategorysystem.com/).

Versioning

This package is versioned independently of ttucs and bumps only when the weights change. ttucs declares a compatible range (ttucs-model-bow>=0.2.0,<0.3). The model's config.json records min_ttucs_version, which ttucs checks at load time.

Changelog

0.2.0 - 2026-11-07

• Retrained with a per-CatID everyday phrase bank, object routing to specific subcategories, 300 examples per CatID and 256-dim embeddings (131,072 × 256, ~68 MB). Hand-written realistic slates: top-1 0.488 → 0.713, top-5 0.813 → 0.925. Still synthetic + UCS data only.

0.1.1 - 2026-11-06

• First published weights: synthetic + UCS data only, compact (131,072 × 128), fp16 safetensors, SHA-256 manifest.