ttucs v0.3.0

Added

• src/ttucs/resources/catid_phrases.yaml: hand-written everyday phrases for 494 CatIDs, used by the
  synthetic generator with probability synthetic.phrase_prob (default 0.5).
• Object routing in DescriptionGenerator: objects named in a specific subcategory's Explanation or
  Synonyms go to that subcategory, and MISC/GENERAL classes only get the leftover objects.
• examples/dev_descriptions.jsonl: 332 hand-written descriptions for choosing settings, so the
  realistic slates stay a held-out test set.
• ttucs eval --predictions writes predictions for the first --data file when no realistic slates
  are evaluated.

Changed

• Default weights are now ttucs-model-bow 0.2.0 (131,072 x 256, ~68 MB, 300 examples per CatID).
  On the 80 hand-written realistic slates, top-1 goes from 0.488 to 0.713 and top-5 from 0.813 to
  0.925; on the dev set, top-1 goes from 0.491 to 0.789. The dependency range is now
  ttucs-model-bow>=0.2.0,<0.3.

Fixed

• Synthetic explanation fragments no longer include cross-references to other classes ("see ...",
  "not ...").

ttucs v0.2.0

Added

• Pretrained weights ship as the data-only ttucs-model-bow 0.1.1 distribution (fp16 safetensors, SHA-256
  manifest, 131,072 x 128 compact model from configs/bow-release.yaml). It is a default dependency,
  so pip install ttucs works offline right after install. It is versioned separately from ttucs.
• ttucs.hub: offline-first model resolution (explicit path, installed package, local cache in
  TTUCS_HOME or the platform cache dir, then a SHA-256-verified GitHub Release download).
  TTUCS_OFFLINE=1 or HF_HUB_OFFLINE=1 disables all network access.
• CLI ttucs info [--verify] and ttucs download [--url --sha256 --force].
• Model loading from weights.safetensors (fp16 or fp32) as well as weights.pt; training records
  its data sources in config.json.
• scripts/build_model_package.py (reproducible weights tarball and wheel), trusted-publishing
  workflow .github/workflows/publish.yml, and docs/RELEASING.md.

Changed

• CI workflows moved from ci/ to .github/workflows/ (ci.yml, publish.yml), so GitHub runs them.
• Lean core install: scikit-learn, pandas, matplotlib and MLflow moved to the [train] extra.
  train, eval and report print an install hint without it.
• TTUCS.from_pretrained() with no arguments now loads the installed ttucs-model-bow weights.
• The default model is now the compact model trained on redistributable data only. Accuracy on the 80
  hand-written evaluation slates drops from 0.713 to 0.488 top-1 (0.913 to 0.813 top-5); synthetic
  test sets stay at about 0.86-0.87 top-1.
• README rewritten for PyPI with absolute links; model and data cards updated for the published model.

Removed

• Breaking: third-party sound library metadata ingestion and its CLI fetch command. The published
  weights are trained only on synthetic data and the UCS list.

Fixed

• License metadata uses an SPDX expression (license = "MIT", license-files) instead of the
  deprecated TOML table and license classifier.

ttucs-model-bow 0.2.0 weights

Release weights for ttucs-model-bow 0.2.0 (fp16 safetensors, 131,072 x 256, synthetic + UCS data only).

Built from commit 6a81c0c (PR #6). Download fallback for ttucs download in ttucs 0.3.0.

sha256: 22e8ac483fa97bbd8152d29bffad60cda668653c7786f8c20e92f90f52913e7b model-bow-v0.2.0.tar.gz

Hand-written realistic slates (n=80): top-1 0.488 -> 0.713, top-5 0.813 -> 0.925. See MODEL_CARD.md for caveats.

ttucs-model-bow 0.1.1 weights

Release weights (fp16 safetensors)