maia-training-data / README.md
ericrisco's picture
Card: link the training process
887ad58 verified
|
Raw History Blame Contribute Delete
2.81 kB
metadata
language:
  - ca
license: other
task_categories:
  - text-generation
tags:
  - andorra
  - catalan
  - sft
  - maia
pretty_name: maia training data
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    data_files:
      - split: train
        path: train.jsonl
      - split: validation
        path: valid.jsonl

maia training data

Conversations in Andorran Catalan used to fine-tune maia, a Gemma 4 model specialised in Andorra. This version is the data of training run 20261104-0036-12b-k-canon (variant K-canon), except for 43 multiple-choice conversations that were trained on but are not published: a wrong option names a person who could not be confirmed as a public figure (the assistant turn is only a letter, and the loss is computed on it alone).

Split Conversations
train 30245
validation 732

Each row is {"messages": [...]} in chat format. Every conversation starts with the same system message, which the model also expects at inference time:

Ets maia, un assistent que coneix a fons Andorra: la seva història, les institucions, la geografia, la cultura i la gent. Respons en català, amb naturalitat i amb la parla d'Andorra.

What is in it

  • Knowledge conversations written from curated articles about Andorra (history, institutions, territory, society, culture, economy, sports, language).
  • An augmented canon of facts: verified atomic facts, each asked about in about 11 different ways (direct, reverse, paraphrase, context, two-turn, false premise, abstention), plus contrast examples between facts that are easy to confuse.
  • Multiple-choice conversations about the same facts, generated by Gemini, whose assistant turn is only the letter of the correct option. The wrong options are invented on purpose, so they are not facts. Do not read them as statements about the people or things they name.
  • probe.jsonl: 338 held-out questions (never in train) with the keywords a correct answer must contain. They measure fact recall before and after training.

The articles reserved by the maia-bench benchmark are excluded, and so are its questions and near-copies of them (word overlap of 0.5 or more). Practice questions can still resemble a benchmark question because both ask about the same facts, and that inflates benchmark scores. See the evaluation in the model card.

Sources and licence

Every source has a provenance record (owner, licence, terms, redistribution status) in the maia repository (docs/fonts/). Redistribution is decided by the project owner per source; see the records before reusing a portion of the data. Code: github.com/ericrisco/maia.

How it was built, including what failed: how maia was trained.