# Attribution and licensing

juurio has three separately licensed parts. The model checkpoints and data files are not bundled in the Python wheel. `juurio setup` downloads them from https://huggingface.co/samuli/juurio and caches them locally.

| part | files | licence |
|---|---|---|
| code | the `juurio` Python package (`juurio/`, the wheel) | MIT, see [`LICENSE`](LICENSE) |
| models | `checkpoints/` (weights, tokenizer and vocabularies), `data/char_vocab.json` | CC BY 4.0, see [`LICENSE-CC-BY-4.0`](LICENSE-CC-BY-4.0) |
| data derived from Giella `fin` | `data/wordlist.json`, `data/fst_table.jsonl`, `data/fst_tag_vocab.json` | GPL-3.0, see [`LICENSE-GPL-3.0`](LICENSE-GPL-3.0) |

For CC BY 4.0 attribution, credit "juurio by samuli, https://huggingface.co/samuli/juurio".

This file is a good-faith summary, not legal advice. Check the upstream licences linked below for your own use.

## Data derived from Giella `fin` (GPL-3.0)

These files are generated by running the Giella `fin` finite-state transducer (through [uralicNLP](https://github.com/mikahama/uralicNLP)) over word forms and lemmas found in a Finnish web corpus, and storing its output. They are derivative works of [GiellaLT `lang-fin`](https://github.com/giellalt/lang-fin) and are distributed under GPL-3.0, the licence of `lang-fin`.

| file | what |
|---|---|
| `data/wordlist.json` | exception table: the inflected forms the neural realizers get wrong |
| `data/fst_table.jsonl` | morphology tags for the 100,000 most frequent word forms where the neural analyser disagrees with Giella (e-mail addresses, @handles, URLs, phone numbers and hashtags removed) |
| `data/fst_tag_vocab.json` | morphology tag vocabulary |

© 2004–present UiT The Arctic University of Norway / Giellatekno, and the `lang-fin` contributors. Source: https://github.com/giellalt/lang-fin (GPL-3.0). `lang-fin` derives from [Omorfi](https://github.com/flammie/omorfi) (GPL-3.0).

## Models (CC BY 4.0)

| file | trained on |
|---|---|
| `checkpoints/realizer.pt`, `verbform.pt`, `finiteverb.pt` | inflected forms generated with Giella `fin` (GPL-3.0) |
| `checkpoints/parser.pt` | [UD_Finnish-TDT](https://universaldependencies.org/treebanks/fi_tdt/index.html) (CC BY-SA 4.0), silver parses of Finnish web text made with [Stanza](https://stanfordnlp.github.io/stanza/) `fi` (Apache-2.0, its model trained on UD_Finnish-TDT), and Giella `fin` morphology tags as input |
| `checkpoints/analyze.pt` | Giella `fin` morphology tags for word forms from Finnish web text |

Whether trained weights are a derivative work of their training data is legally unsettled. The table documents where the training data came from, so that users can make their own judgement, and credits the upstream sources.

## Tools used to build the models and data

None of these are needed to run juurio.

| project | use | licence |
|---|---|---|
| [GiellaLT `lang-fin`](https://github.com/giellalt/lang-fin) through [uralicNLP](https://github.com/mikahama/uralicNLP) | the morphology the realizers learn from | GPL-3.0 |
| [Stanza](https://stanfordnlp.github.io/stanza/) | silver dependency parses | Apache-2.0 (code) |
| [UD_Finnish-TDT](https://universaldependencies.org/treebanks/fi_tdt/index.html) | parser training data, and the treebank behind Stanza's `fi` model | CC BY-SA 4.0 |
| [Voikko](https://voikko.puimula.org/) | checking generated forms | MPL-1.1 / GPL-2.0+ / LGPL-2.1+ (Finnish dictionary GPL) |
| [Omorfi](https://github.com/flammie/omorfi) | upstream of `lang-fin` | GPL-3.0 |

## Runtime dependencies

`torch` (BSD-3-Clause), `transformers` (Apache-2.0), `numpy` (BSD-3-Clause), `huggingface-hub` (Apache-2.0).