softcatala/mantinc-catalan-drift

Mantinc — Catalan Drift Benchmark

Descripció (ca)

Mantinc és un banc de proves que avalua si un model de llenguatge continua responent en català quan el missatge, la conversa prèvia o el context recuperat l'empenyen a fer-ho en una altra llengua, normalment el castellà o l'anglès.

Dataset Description

Mantinc is a benchmark that measures whether a language model keeps answering in Catalan when the prompt, prior conversation, or retrieved context tries to pull it into another language (typically Spanish or English).

It targets the language-confusion problem studied by Marchisio et al. in "Understanding and Mitigating Language Confusion in LLMs" and extends it with Catalan-specific personas, workflows, and multi-turn pressure cases.

• Source repository: https://github.com/Softcatala/mantinc-catalan-drift
• Benchmark design and scorer calibration: see benchmark_design.md in the source repository.
• lm-eval task definition: lm_eval_tasks/catalan_drift/ in the source repository.

Scope

What this benchmark measures: whether the model's final answer is in Catalan despite source-language pressure from prompts, prior turns, or retrieved context. Scoring uses a fastText language check over response segments.

What this benchmark does not measure: task completion, factual accuracy, tone, formatting, or overall answer quality. A model that refuses, goes off-topic, or produces a poor answer in fluent Catalan will still pass. A high score means "keeps the language," not "is a good Catalan assistant."

Dataset

The 300-item test split is balanced across five scenario categories:

• monolingual — Catalan-only inputs, no cross-lingual pressure
• crosslingual_basic — Catalan prompt with limited other-language content
• multi_turn — Multi-turn conversations that gradually shift language
• crosslingual_advanced — Heavier cross-lingual mixing (templates, quoted spans, technical content in ES/EN)
• rag_context — Retrieved context in Spanish or English, prompt in Catalan

Fields

Each row is a JSON object with:

Field	Type	Description
id	string	Stable item identifier (e.g. cd_001)
category	string	One of the five categories above
persona	string	Persona label used when authoring the prompt
workflow	string	Task/workflow archetype
source_lang	string	Language of the surrounding context
target_lang	string	Expected answer language (always ca)
prompt	string	The full model-visible input (may include conversation and retrieved context)

Usage

With datasets

from datasets import load_dataset

ds = load_dataset("softcatala/mantinc-catalan-drift", split="test")
print(ds[0])

With lm-eval

The benchmark is designed to run as an lm-eval-harness task. Task definition lives in the upstream repository at lm_eval_tasks/catalan_drift/. See the source repository for the full run instructions, including how to install the fastText language identifier used for scoring.

Scoring

Segment-level Catalan pass rate. Each response is split into segments, each segment is classified by fastText (lid.176), and an item passes iff the non-Catalan token ratio stays under 15%. Thresholds are calibrated against a FLORES-200 slice — see benchmark_design.md in the source repository.

Reference results

Numbers below are point estimates on the 300-item main set, conditioned on the decoding config each row was evaluated under. They are not directly comparable across rows unless the configs match, and small rank differences inside the confidence intervals should not be read as capability gaps.

Model	Overall	Catalan token ratio
Gemma 4 26B A4B	82.7% ±4.3	84.9%
Ministral 3 8B	82.3% ±4.3	72.0%
ALIA 7B	81.3% ±4.4	71.8%
Gemma 4 12B	79.3% ±4.6	81.0%
Ministral 3 14B	78.0% ±4.7	80.3%
Gemma 3 12B	75.0% ±4.9	79.0%
Gemma 4 E4B	73.3% ±5.0	82.0%
GPT-5.6	71.3% ±5.1	76.2%

± N is the Wilson 95% half-width at n=300.

License

The dataset (prompts, fixtures, evaluation data) is released under the Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0) license. The associated code in the upstream repository is licensed under MIT.

Citation

If you use Mantinc in your research, please cite the upstream repository:

@misc{mantinc2026,
  title  = {Mantinc: A Catalan Language-Drift Benchmark},
  author = {Jordi Mas i Hernàndez, Softcatalà},
  year   = {2026},
  url    = {https://github.com/Softcatala/mantinc-catalan-drift}
}

Contributing

The dataset needs review from real-world cases: checking that prompts represent plausible situations, proposing new cases, and running evaluations with other models. Contributions are welcome in the upstream repository.