# NOTICE

GameDecider v0.1 (`rel5_ettin400_mixI_B_safe3_soup_enc`)
Copyright (c) 2026 Holdout Labs

The model weights, adapters, sample rows and the `runtime/` package (gamedecider 0.1.0) are released under the Apache
License 2.0 (see `LICENSE`).
This file lists third-party material the weights are derived from, and the credits the dataset licences ask for.
Row counts and the licence evidence per source are in `DATA_SOURCES.md` / `data_sources.json`.

## 1. Base model: Ettin encoder 400M (MIT)

GameDecider is fine-tuned from `jhu-clsp/ettin-encoder-400m`, revision `7662476d60abb071a5bd319c9f3074f3072c062d`
(https://huggingface.co/jhu-clsp/ettin-encoder-400m), licensed MIT. The tokenizer (`tokenizer.json`) is Ettin's, same
revision, unchanged.

```
MIT License

Copyright (c) 2025 Johns Hopkins University Center for Language and Speech Processing

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.
```

## 2. Datasets under Creative Commons Attribution 4.0 (CC BY 4.0)

Licence: https://creativecommons.org/licenses/by/4.0/ . All of these were **modified**: dialogue and text were cut into
game-style states, sometimes renamed, and recast into decision questions with new question wording. The original
datasets are not redistributed with the weights.

| Dataset | Credit | Link |
|---|---|---|
| MASSIVE | FitzGerald et al. 2022, Amazon.com, Inc. | https://huggingface.co/datasets/AmazonScience/massive |
| HateCheck | Röttger et al. 2021 | https://huggingface.co/datasets/Paul/hatecheck |
| Banking77 | Casanueva et al. 2020, PolyAI | https://huggingface.co/datasets/PolyAI/banking77 |
| CaSiNo | Chawla et al. 2021 | https://huggingface.co/datasets/kchawla123/casino |
| HWU64 (NLU-Evaluation-Data) | Liu et al. 2019 | https://github.com/xliuhw/NLU-Evaluation-Data |
| Taskmaster-2 | Byrne et al. 2020, Google LLC | https://github.com/google-research-datasets/Taskmaster/tree/master/TM-2-2020 |
| Diplomacy deception ("It Takes Two to Lie") | Peskov, Cheng, Elgohary, Barrow, Danescu-Niculescu-Mizil and Boyd-Graber, ACL 2020 | https://github.com/DenisPeskoff/2020_acl_diplomacy |
| FIREBALL | Zhu, Aggarwal, Feng, Martin and Callison-Burch, ACL 2023 (revision 1b363686) | https://huggingface.co/datasets/lara-martin/FIREBALL |
| Circa | Louis et al. 2020, Google Research | https://huggingface.co/datasets/google-research-datasets/circa |
| IRC disentanglement | Kummerfeld et al. 2019, "A Large-Scale Corpus for Conversation Disentanglement" | https://huggingface.co/datasets/jkkummerfeld/irc_disentangle |
| AMI Meeting Corpus (manual annotations 1.6.2) | Carletta et al. 2005, AMI Consortium | https://groups.inf.ed.ac.uk/ami/corpus/license.shtml |
| ATOMIC 2019 | Sap et al. 2019, Allen Institute for AI | https://huggingface.co/datasets/allenai/atomic |
| DynaHate (Dynamically Generated Hate Speech) | Vidgen et al. 2021, "Learning from the Worst" | https://github.com/bvidgen/Dynamically-Generated-Hate-Speech-Dataset |

## 3. Datasets under the Open Data Commons Attribution License (ODC-By 1.0)

Licence: https://opendatacommons.org/licenses/by/1-0/

- **CODAH**, Chen et al. 2019, https://github.com/Websail-NU/CODAH. Modified (recast into
  decision questions).

## 4. Datasets under MIT

The MIT licence text in section 1 applies, with these copyright lines:

- **CICERO redacted games** (`data/cicero_redacted_games`), Meta Fundamental AI Research Diplomacy Team (FAIR) et al.,
  Science 2022, https://github.com/facebookresearch/diplomacy_cicero. Copyright (c) Meta, Inc. and its affiliates.
- **MultiLIGHT** (multi-party LIGHT, ParlAI), Wei et al. 2023, https://parl.ai/projects/light/ ,
  http://parl.ai/downloads/projects/multilight/parlai_multilight.tar.gz . Copyright (c) Facebook, Inc. and its affiliates.
- **IGLU** clarifying-questions data, Mohanty et al. 2023, https://github.com/microsoft/iglu-datasets. Copyright (c) Microsoft Corporation. MIT License.

## 5. Datasets under Apache-2.0

Licence: https://www.apache.org/licenses/LICENSE-2.0

- **GoEmotions**, Demszky et al. 2020, Google, https://huggingface.co/datasets/google-research-datasets/go_emotions
- **deepset prompt-injections**, deepset GmbH, https://huggingface.co/datasets/deepset/prompt-injections
- **PersuasionForGood**, Wang, Shi, Kim, Oh, Yang, Zhang and Yu, ACL 2019, https://gitlab.com/ucdavisnlp/persuasionforgood

## 6. Public domain and CC0 (credit given as a courtesy)

- **Civil Comments**, Borkan et al. 2019 (Jigsaw), CC0-1.0, https://huggingface.co/datasets/google/civil_comments
- **Snips NLU benchmark** and **Snips built-in intents** (Sonos), CC0-1.0, https://github.com/sonos/nlu-benchmark ,
  https://huggingface.co/datasets/sonos-nlu-benchmark/snips_built_in_intents
- **80 public-domain novels and plays** from Project Gutenberg (US: published 1930 or earlier; Canada: author and
  translator died 1971 or earlier). Characters were renamed. The list is in `data_sources.json`.

## 7. Synthetic data from open-weight models

About half of the training rows were written or labelled by open-weight models run through commercial inference
providers (Novita, DeepInfra, Baseten). The models and their licences:

- GLM-5.3-Flash (Z.ai), MIT
- DeepSeek-V4.1-Flash and DeepSeek-V4.1-Flash-1017 (DeepSeek), MIT
- gpt-oss-120b (OpenAI, open weights), Apache-2.0 (labeller and checker only)
- Qwen3.8-27B and Qwen3.6-35B-A3B (Alibaba Qwen), Apache-2.0 (task-family catalogues only)

These licences don't require attribution for outputs. They are listed for transparency.