symphonym-v8 / LICENCE.md
docuracy's picture
Upload Symphonym v8 model, vocabularies, and evaluation results
c02bc8b verified
|
Raw History Blame Contribute Delete
7.83 kB
# Licensing
**This repository is not one licensable thing**, and a single grant at the root would
misdescribe most of it. What follows states what is settled, what is not, and one
exclusion that must never become untrue.
## Component map
| component | status |
|---|---|
| **`zenodo/epitran_extensions/`** β€” graphemeβ†’IPA rule sets | βœ… **Contributions under CC0 1.0.** WHG proceeds on the basis that such rows attract no rights, and the CC0 dedication exists so nothing turns on it. ⚠ The v7 Zenodo record states CC BY 4.0 at the record level; that stands for v7 and is to be revisited for v8. See [`zenodo/epitran_extensions/LICENCE.md`](zenodo/epitran_extensions/LICENCE.md). |
| **Model weights** (`hf/model.safetensors`) | βœ… **CC BY 4.0**, as published within the Zenodo deposit (DOI [10.5281/zenodo.18682017](https://doi.org/10.5281/zenodo.18682017), 2026-02-22). βœ… Checked 2026-09-07: `hf/README.md` already declares `license: cc-by-4.0`, so the model card and the deposit agree. |
| **WHG's own code** | βœ… **BSD 3-Clause** β€” see [`LICENSE`](LICENSE). Matches `WorldHistoricalGazetteer/whg3`, same copyright holder (University of Pittsburgh), same scoping to "WHG's own code". |
| **Third-party dependencies** | Unchanged, under their own terms. Nothing here alters them. |
| **Ingested authority data** | Not licensed by WHG; see below. |
## Ingested authority data is not ours to license
`processing/settings.py` records a `license_spdx` per authority. Twelve distinct terms
are currently declared, including `CC0-1.0`, `CC-BY-4.0`, `CC-BY-SA-4.0`,
`ODbL-1.0`, `ODC-By-1.0` and four bespoke `custom-*` terms β€” **and `CC-BY-NC-4.0` and
`CC-BY-ND-4.0`.**
⚠ **The last two are named rather than smoothed over.** This project's outputs are
derivatives by construction: toponyms are extracted, transcribed and embedded. An
**ND** source is therefore a genuine unresolved question, not a labelling detail, and
a **NC** source constrains commercial reuse of anything derived from it. **This file
does not resolve either**, and no statement here should be read as a grant over data
WHG did not create.
Per-source terms are surfaced to consumers through the gazetteer registry and the
`namespaces[]` field of API responses. **Ingestion is deliberately not gated on
licence** β€” attribution is handled at the point of use.
## πŸ›‘ EXCLUDED CONTENT β€” must never enter this repository
Some material used by this project is cleared for **acquisition and research use
only**, and **not for publication or redistribution**. It is held outside the
repository and must stay there.
**Currently excluded:** the RCAHMW Long-Held Place Names (LHPN) Welsh/English
name-variant data, held at `/vast/ishi/lhpn` with its own licence notice. Cleared
verbally for acquisition and model training; **publication and redistribution are NOT
cleared.**
⚠ **Nothing in this file grants any licence over excluded content, and no permissive
statement here should be read as a reason to commit it.**
βœ… **This is enforced mechanically, not by prose.** `tests/test_no_restricted_content.py`
fails on any path or file content matching the restricted set. **An exclusion that
only exists in a document is honoured until the first person who has not read it** β€”
if you are adding restricted material, that test is what should stop you, and if it
does not, the test is the defect.
## Attribution
> World Historical Gazetteer, *Symphonym: Universal Phonetic Embeddings for
> Cross-Script Name Matching β€” Models and Evaluation Data*. Zenodo.
> https://doi.org/10.5281/zenodo.18682017
## Required of the v8 deposit
βœ… **Decided (SG, 7 Sep 2026): the relationship between the record-level licence and
the per-component terms must be STATED at the deposit for v8, not left to be
inferred.**
The v7 record states **CC BY 4.0** at record level while the rule sets inside it are
dedicated **CC0**. CC0 is the more permissive, so a recipient relying on the
record-level licence is never *wrong* β€” they would simply be over-complying for part
of the deposit. That is an ambiguity rather than a conflict, and it is resolved by
saying so.
πŸ›‘ **Concretely, and this closes the defect that prompted the whole exercise: SHIP THE
LICENCE FILES INSIDE THE DEPOSIT.** The v7 archive contains **no `LICENSE`, `COPYING`
or `NOTICE` anywhere** β€” the CC BY 4.0 claim exists *only* in Zenodo's record
metadata, so anyone who takes the files from the archive, or from GitHub, sees no
licence at all. ⚠ **A licence that lives only in a catalogue record does not travel
with the bytes.**
So the v8 deposit must:
1. **Contain `LICENCE.md` and `zenodo/epitran_extensions/LICENCE.md`**, so the terms
travel with the files.
2. **State in the record description that the record-level licence is a floor**, with
per-component terms given inside β€” rather than implying one licence governs
everything in the deposit.
3. **Attribute the upstream sources, not only WHG.** ⚠ The attribution obligations in
the deposit are **inherited**: GeoNames (**CC BY 4.0**) and Getty TGN (**ODC-By
1.0**) both require attributing *them*. **A record saying "CC BY 4.0, attribute
WHG" understates the obligation.** Name GeoNames and Getty TGN as upstream sources
β€” which is also the house position everywhere else: WHG asserts its terms
*alongside* each source's, never instead of them.
βœ… **The "floor" framing was checked before being relied on, and the check is the
generalisable part.** *"The record-level licence is a floor"* is only safe if **nothing
in the deposit is MORE restrictive than the record** β€” otherwise a recipient relying on
it **over-claims rights**, which is a worse failure than over-complying. Verified
against the training inputs: GeoNames **CC BY 4.0**, Wikidata **CC0**, Getty TGN
**ODC-By 1.0** β€” **no share-alike and no NonCommercial**, so the most restrictive input
is attribution-required and the record-level CC BY 4.0 matches it. ⚠ **This could
easily have gone the other way**: WHG's index as a whole is ~42.7% share-alike,
dominated by OSM/ODbL, and an ODbL or NC source in the *training set* would have raised
a real question. Those sources are not in it.
## Checked and found not to apply
**NEH funding.** βœ… **All of the work in this repository postdates the NEH funding, so
NEH's terms do not reach any of it** (SG, 7 Sep 2026). Recorded rather than left
silent, because the question will be asked again and re-deriving the answer costs more
than reading it.
⚠ **What the terms would have required, had they applied, and one part of it is not
what people expect.** NEH General Terms Β§XI.B–C (2 CFR Β§200.315–316) leave intangible
property with the recipient, while the Federal Government retains a **non-exclusive,
royalty-free, irrevocable** licence to reproduce, prepare derivative works from,
publish and use it, and to authorise others to do so. Open licensing is a *superset*
of that, so it would have constrained nothing here β€” **but because the reservation is
irrevocable and non-exclusive, it would block an EXCLUSIVE commercial licence over
anything it did cover.** πŸ›‘ **That is a live consideration for WHG generally, even
though it does not touch this repository**, and it is the half most likely to be
overlooked, since the reservation reads as permissive until someone tries to grant
exclusivity.
## Resolved elsewhere
βœ… `whg3`'s `NOTICE.md` previously read as applying **CC BY-NC 4.0** to contributed datasets. ⚠ **The error was scope, not licence** β€” NC is WHG's *curation overlay*, asserted alongside each source's own terms, and applying it to other people's material was self-defeating besides, since CC BY 4.0 Β§2(a)(5)(B) forbids imposing downstream restrictions on material WHG received under it. Corrected in `whg3` (`080bc72c4`).