Tacet
Copyright 2026 CodePawl

This product is licensed under the Apache License, Version 2.0 (see LICENSE).
It includes or builds on the third party work listed below.


Software
========

Laya (https://huggingface.co/convaiinnovations/laya, pip package laya 0.3.5)
Copyright Convai Innovations. Licensed under the Apache License, Version 2.0.
Tacet does not depend on the laya package. The following parts are copied from
laya/common.py and adapted (renamed, split into smaller functions):
  src/tacet/packing.py   question type ids, criterion, option and state rendering
                         (the prompt format the model was trained on)
  src/tacet/network.py   layout of the two layer head and the option scorer
  src/tacet/decoding.py  normalized entropy confidence
  scripts/benchmark.py   15 bin expected calibration error, and the scoring
                         definitions from Laya's typed-decisions evaluation notebook
The packed sequence (all questions and the state in one pass) and the per token
segment embedding are CodePawl's own.


Model weights (codepawl/tacet-sonata on Hugging Face)
===============================================

The released weights are fine tuned from mmBERT:

mmBERT (jhu-clsp/mmBERT-small)
By the Johns Hopkins University Center for Language and Speech Processing (JHU CLSP).
Licensed under the MIT License:

  Permission is hereby granted, free of charge, to any person obtaining a copy
  of this software and associated documentation files (the "Software"), to deal
  in the Software without restriction, including without limitation the rights
  to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
  copies of the Software, and to permit persons to whom the Software is
  furnished to do so, subject to the following conditions:

  The above copyright notice and this permission notice shall be included in all
  copies or substantial portions of the Software.

  THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  SOFTWARE.

The mmBERT Hugging Face repositories declare the MIT License and ship no
separate copyright line, so the notice above names the authors as listed there.
The tokenizer files in the released models are mmBERT's tokenizer files, which
mmBERT distributes under the same MIT licence; mmBERT's documentation says its
vocabulary is the Gemma 2 tokenizer.


Training data
=============

  LocalLLaMA/typed-decisions            Apache License 2.0 (train split only;
                                        the test split is held out for evaluation)
  AmazonScience/massive (MASSIVE)       CC BY 4.0
                                        Copyright Amazon.com, Inc.
  PolyAI/banking77                      CC BY 4.0
                                        Copyright PolyAI
  bitext/Bitext-customer-support-llm-chatbot-training-dataset
                                        Community Data License Agreement,
                                        Sharing, Version 1.0. Under its section
                                        3.5, the licence places no obligations
                                        on Results (the trained model).
  google/civil_comments                 CC0 1.0
  Synthetic cases and labels generated with Qwen3-30B-A3B-Instruct-2507 and
  Qwen3-30B-A3B-Thinking-2507 (Apache License 2.0).

Not used: clapAI/MultiLingualSentiment (upstream licences unknown), any share
alike dataset, and any text written by a proprietary model.

The CC BY 4.0 datasets were adapted into typed decision cases (questions and
answer distributions were derived from their labels).
