Champollion-LYSS

Community language-validation plugins for the MT Eval Harness.

Part of Champollion — open-source infrastructure for trustworthy machine translation across every language, built with communities and never scraped from them. The infrastructure is open-source and singly stewarded; the test sets and the methods for a community's language belong to that community. LYSS is how a language community's own validation standard travels into the harness on its own terms. Explore the wider network at champollion.dev/docs/network.

LYSS plugins are the referee of MT evaluation: FST-gated linters and semantic validators that judge whether a translation is morphologically valid and meaning-preserving in the target language — signals that surface metrics (chrF++, BLEU) cannot see. They live with the language, not with any translation method, so they score every method fairly.

The harness stays strictly language-neutral: it ships no language-specific scorer code. Instead it fetches and loads the relevant LYSS plugin on demand, driven by a language card's evalMetrics / evalStandard declaration.

This repo is the home for all such plugins. Today:

Install

pip install champollion-lyss            # core (linters)
pip install "champollion-lyss[crk]"     # + spaCy for the Cree semantic validator

The harness installs this automatically when an evaluation targets a language whose card declares a LYSS standard (mt-eval setup --lang crk).

Data sovereignty & responsible use

This package ships code only — it bundles no dictionary, gloss, lexicon, or corpus data. The Plains Cree validation standard and any Cree language data are treated as the property of the nêhiyaw (Plains Cree) language community and are offered for non-commercial, community-benefit use, aligned with Indigenous data sovereignty principles (OCAP®). Ownership is split deliberately: Champollion maintains at most the GENERIC evaluation methodology (the LYSS standard); the Cree instantiation — the linguistic knowledge it validates against — is not ours, and models or artifacts built for Cree are governed by the community (see the Derived-Artifacts Commitment on champollion.dev). Commercial use of the Cree standard is reserved pending community consent. The full statement is in the NOTICE file, which travels with every redistribution.

Language-specific resources are obtained at runtime — fetched, never scraped or bundled:

• Glosses → itwêwina API. The Cree semantic validator fetches glosses per-lemma from the public itwêwina API (itwewina.altlab.app) and caches them under ~/.mt-eval/itwewina-cache/. The underlying dictionary content (Wolvengrey CW, Maskwacîs MD, AECD) is not openly licensed: the cache is for your own local use only — never commit, redistribute, publish, or bundle it.
• IP acknowledgment. Before any gloss is fetched you are shown the terms and must acknowledge them. In automation set CHAMPOLLION_LYSS_ACCEPT_IP=1 (CI / harness auto-setup acknowledge automatically; the notice is always printed).
• FST → downloaded, invoked as a separate tool. Morphological analysis uses a finite-state transducer (e.g. the GiellaLT/ALTLab Plains Cree FST, AGPL) that the harness fetches on demand to ~/.mt-eval/fsts/crk/. It is never bundled into or linked into this package.
• Fail honestly. If the FST, the spaCy model, or the gloss API is unavailable (or IP terms are unacknowledged), the affected score is reported as unavailable — never fabricated or substituted with synthetic data (human-validated data only).
• Inline grammar facts are cited, not extracted. The small hand-curated lists in the code (optional particles, lemma-synonym pairs, structural words) are widely-published, cross-referenced facts of Plains Cree grammar — citable to multiple published grammars and lexica, references to published fact, not bulk-extracted data.

License

Use by permission only (interim; SPDX LicenseRef-Champollion-Interim-Permission-Required, see LICENSE). The Plains Cree standard is built on ALTLab/GiellaLT infrastructure and published Cree scholarship, and it encodes decisions that properly belong with the nêhiyaw language community. Until those consultations have happened, we take the conservative posture: written permission is required for any use — and is ordinarily granted without charge for noncommercial research, education, and community-benefit use. No commercial permission is granted at all, pending community consent. Source-available, not OSI "open source", and deliberately so for now.

In code terms the package stays independent of the data it fetches and the tools it invokes: the GiellaLT/ALTLab Cree FST (AGPL) is invoked as a separate downloaded tool, never bundled or linked, and no licensed dictionary/corpus data is shipped. The permission requirement is a stance about authority over the standard, not a claim that this code embeds ALTLab's.

The MT-Eval Harness this plugs into is AGPL-3.0, and grants an additional permission under AGPL §7 allowing the harness to combine with separately-licensed eval-standard plugins like this one through its public plugin interface — so a permission-gated plugin and the AGPL harness compose cleanly.

Monorepo note

Development happens in the Champollion monorepo (lyss/). This standalone repository exists purely as a license boundary — the interim permission-only license must not blur into the monorepo's Apache/AGPL code — and receives deliberate snapshot updates from the monorepo (there is no automatic mirroring).

OCAP® is a registered trademark of the First Nations Information Governance Centre (FNIGC): https://fnigc.ca/ocap-training/