Open source is language infrastructure

When a language depends on one company's private model, it also depends on that company's pricing, priorities, moderation policy, product roadmap and willingness to keep supporting the language.

That is fragile infrastructure. For Irish and other lower-resource languages, openness is not a philosophical extra. It determines whether a university can reproduce a result, whether a public body can audit a system, whether a small company can adapt it, and whether speakers can challenge how their language is represented.

Open weights are useful, but they are not the whole source

AI uses the word open loosely. A downloadable checkpoint can be valuable even when its licence restricts use or its training process cannot be reproduced. It should still be described accurately.

The Open Source Initiative's Open Source AI Definition 1.0 draws a clearer line. The preferred form for modifying a machine-learning system includes three broad elements:

• sufficient information about the training data, including provenance, selection, labelling and processing;
• the code used to process data, train, validate, test and run the system;
• the model parameters under terms that permit open use.

This does not mean every training row can always be published. Privacy, copyright, contractual limits and speaker consent can prevent release. It does mean the absence should be documented well enough for others to understand what shaped the model.

Why this matters more when data is scarce

English model development benefits from an enormous supply of text, code, benchmarks, researchers and commercial demand. A mistake in one dataset can be diluted, discovered or corrected by many independent efforts. A lower-resource language has less redundancy.

A small corpus may carry disproportionate influence. A narrow source can make a model sound formal in every setting. Duplicated material can inflate benchmark results. Machine translation can multiply the same error across a synthetic dataset. A model can appear fluent to a non-speaker while failing on grammar, dialect or ordinary idiom.

Releasing only a score conceals those problems. Releasing the evaluation prompts, judgements, split logic, data notes and model card gives the community something it can test and improve.

In language AI, reproducibility is a form of cultural resilience.

Open work lets communities compound progress

The strongest examples do more than upload weights. BigScience built BLOOM through an international open research collaboration and released a detailed model card describing intended use, limitations and training context. Its ROOTS corpus work documented a 1.6 TB collection spanning 59 languages, along with data-preparation tools and a release strategy shaped by the rights attached to each source.

Masakhane offers an even more relevant lesson for lower-resource language work. Its participatory machine-translation research treated low-resourcedness as a social problem, not simply a missing-data problem. Speakers and local researchers contributed datasets, benchmarks, human evaluations, models and papers across more than 30 African languages.

The transferable principle is not that Irish should copy either project. It is that people who know a language must be able to shape the research, not arrive at the end to rate a finished system.

Openness creates practical options

It makes independent evaluation possible

A school, broadcaster or public body can test a model against its own language requirements instead of trusting a vendor's multilingual average. Failures can be published as reusable test cases.

It supports local adaptation

Open parameters and training code let a small lab adapt a strong base for Irish legal retrieval, educational feedback, speech recognition or public-service terminology without training a foundation model from zero.

It reduces strategic lock-in

Organisations can host locally, move providers, merge an adapter, or continue maintaining a model after the original publisher changes direction. That matters for services expected to last longer than a product cycle.

It makes public funding travel further

When suitable outputs from publicly supported work are released with clear licences and documentation, the next project can begin from a baseline instead of repeating the same collection and evaluation work.

Open does not mean careless

Publishing a dataset can expose personal information, copyrighted material or speech that contributors never expected to train a model. Publishing a model can also create new misuse risks. A serious open process makes those tensions visible.

For language projects, the release decision should be made artifact by artifact:

• open evaluation code and rubrics wherever possible;
• publish model cards and data statements for every release;
• release datasets only when rights and consent support it;
• provide derived statistics when source text cannot be shared;
• use licences that match the real permissions and restrictions;
• keep sensitive domain adapters private when confidentiality requires it.

Mozilla Common Voice provides a concrete pattern for community speech data. Its Irish scripted-speech release is available under CC0 with a datasheet describing clips, validated hours, speakers and collection rules. That documentation is part of the asset.

Our release standard

LeemerLabs will not call every downloadable model open source. We will name the actual licence and state what is present: weights, adapters, code, data manifests, evaluation sets, results and known limitations.

For Born releases and planned Gaeilge work, our target is a useful release bundle:

1. a model or adapter with an explicit licence;
2. a model card that names the base, task, limits and intended use;
3. training configuration and reproducible commands where publishable;
4. data provenance at source and transformation level;
5. held-out evaluations with answer sheets or judging rubrics;
6. negative results when they change how the evidence should be read.

Fine-tuning services can follow the same discipline even when client data and weights remain private. The client should still receive the training record, provenance, baseline comparison, evaluation harness and exportable artifacts. Ownership without documentation is only partial ownership.

The goal is local capacity

Open language technology is successful when more people can do useful work after the release: speakers can test it, teachers can critique it, researchers can reproduce it, companies can adapt it and institutions can deploy it on terms they understand.

That is the future we want for Irish AI. Not one lab controlling the language layer, but a growing body of shared infrastructure with room for public, academic, community and commercial work.

If you maintain an Irish dataset, evaluation, model or public-interest tool and want to make it easier for others to build on, contact hello@leemerlabs.ie.