ConlangCrafter: Constructing Languages with a Multi-Hop LLM Pipeline

Project Page: conlangcrafter.github.io

Paper: arxiv.org/abs/2508.06094

Dataset: huggingface.co/datasets/malper/ConlangCrafter — 64 generated languages

We introduce a fully automated system for constructing languages (conlangs) using large language models. Our multi-stage pipeline creates coherent, diverse artificial languages with their own phonology, grammar, lexicon, and translation capabilities.

Install dependencies:

pip install -r requirements.txt

Set up API keys— copy.env.example

to.env

and add keys for whichever APIs you will use:Google Gemini:GOOGLE_API_KEY

—Google AI StudioOpenAI:OPENAI_API_KEY

—OpenAI API KeysDeepSeek (via Together):TOGETHER_API_KEY

—Together AI

Generate a language sketch(default model:gemini-2.5-pro

):

python src/run_pipeline.py

Run python src/run_pipeline.py --help

to see all options. Key flags:

python src/run_pipeline.py \
    --model gemini-2.5-pro \
    --custom-constraints "The language has only 3 vowels" \
    --temperature 0.8 \
    --qa-disabled        # QA self-refinement loops are on by default; use this to turn it off

To resume a previous run (e.g. starting from grammar after phonology completed):

python src/run_pipeline.py --language-id <id> --steps grammar,lexicon

Supported models are:

• Google Gemini (e.g., gemini-2.5-pro

,gemini-1.5-flash

) - OpenAI models (e.g., o4-mini

,gpt-4o

,gpt-5

) - DeepSeek via Together AI (e.g., deepseek-ai/DeepSeek-R1

)

You can load pregenerated language sketches from our dataset in this pipeline's format with this script:

python src/load_hf_languages.py

Translation is not run by default. To translate into a generated language, run the translation step separately. By default it translates the 10 sentences in configs/sentences_default.txt

:

python src/run_pipeline.py --language-id <id> --steps translation

To translate a single custom sentence instead:

python src/run_pipeline.py --language-id <id> --steps translation --translation-sentence "Hello, world!"

Pass --translation-sketch-update

to feed new vocabulary and grammar rules introduced during translation back into the sketch for each subsequent sentence, expanding the language as translation proceeds (constructive translation).

This implementation includes minor improvements to the system used for results from our paper:

QA loop: Degenerate outputs (e.g. JSON instead of text) are detected and skipped inline, rather than post-hoc rejection sampling.QA amend prompt: Prompt wording is slightly adjusted for consistency with our system.

@article{conlangcrafter2025,
    title={ConlangCrafter: Constructing Languages with a Multi-Hop LLM Pipeline},
    author={Morris Alper and Moran Yanuka and Raja Giryes and Ga{\v{s}}per Begu{\v{s}}},
    year={2025},
    eprint={2508.06094},
    archivePrefix={arXiv},
    primaryClass={cs.CL},
    url={https://arxiv.org/abs/2508.06094}
}

This project is licensed under the MIT License — see the LICENSE file for details.