Dataset Card for Bilingual-BabyLM

Bilingual-BabyLM dataset collected as part of the paper "Bringing Up a Bilingual BabyLM: Investigating Multilingual Language Acquisition Using Small-Scale Models" by Linda Zeng, Steven Y. Feng, and Michael C. Frank. For more details, please see Appendices A-B in our paper.

The dataset is a synthetic bilingual and code-switching dialogue dataset designed for training and evaluating small language models in multilingual and code-switched settings. It contains parallel English, Spanish, inter-sentential code-switching, and intra-sentential code-switching versions of parent-child dialogues. Dialogues are categorized by speaker identity and conversational context.

Dataset Sources

• Repository: https://github.com/lindazeng979/bilingual-babyLM
• Paper: https://arxiv.org/abs/2603.29552

---

Dataset Structure

The Hugging Face dataset repository contains three primary dataset files:

• bilingual-babyLM.jsonl
  Raw synthetic bilingual dialogue dataset containing parallel English, Spanish, inter-sentential code-switching, and intra-sentential code-switching versions of each dialogue.
• BL_conditions.zip
  Final processed training-condition datasets used in the main experiments (~100M-word training conditions).
• BL_conditions_20M.zip
  Reduced-data 20M-word versions of the training conditions used for low-resource experiments.

Bilingual-BabyLM Dialogues

The raw dialogue dataset contains approximately 800k parallel child-directed dialogues (~100M words per dialogue variant; ~400M total words).

Each row of bilingual-babyLM.jsonl contains the following fields:

• custom_id: Unique identifier for the dialogue example (including number of dialogue turns).
• content_en: English version of the dialogue.
• content_es: Spanish version of the dialogue.
• content_intersent: Inter-sentential code-switching version, where language alternates across sentences or turns.
• content_intrasent: Intra-sentential code-switching version, where English and Spanish may be mixed within the same sentence or turn.
• parent: Parent speaker identity, such as mom or dad.
• context: Dialogue setting, such as home or community.

Example:

{
  "custom_id": "BL_dadcommunity_10-turns_batch9_ex-2124",
  "content_en": "**Dad**: Look over there! Do you see that big, fluffy zebra?...",
  "content_es": "**Dad**: ¡Mira allá! ¿Ves esa cebra grande y esponjosa?...",
  "content_intersent": "**Dad**: Look over there! ¿Ves esa cebra grande y esponjosa?...",
  "content_intrasent": "**Dad**: Look over there! ¿Ves ese zebra grande y peludo?...",
  "parent": "dad",
  "context": "community"
}

The dialogue text includes speaker tags such as **Mom**,**Dad** and **Child**, newline markers, and <|endoftext|> delimiters used for causal language modeling.

Training Condition Folders

The zip archives contain many folders, where each folder corresponds to a specific training condition used in the paper experiments. Each condition folder contains train.txt and val.txt.

The processed training conditions include:

• English-only training (toplines and baselines)
• Spanish-only training
• multilingual bilingual mixtures
  • random bilingual mixing
  • speaker-conditioned bilingual mixing
• inter-sentential code-switching
• intra-sentential code-switching
• varying exposure conditions
• 20M-word reduced-data subsets

Folder names directly correspond to the training conditions described in the paper.

---

Dataset Creation

Curation Rationale

The dataset was created to study multilingual and code-switching language learning in small language models under controlled data conditions.

The goal is to provide:

• parallel bilingual dialogue data across English, Spanish, and code-switching conditions
• dialogue data categorized by speaker identity and conversational context
• synthetic inter-sentential and intra-sentential code-switching data
• controlled data conditions used to test small language models in BabyLM setup

Data Collection and Processing

Building upon the TinyDialogues generation framework, we used GPT-4-based models to synthesize approximately 800k parallel child-directed conversations across English, Spanish, and code-switching conditions. The dialogues additionally vary across child age, conversation type, speaker identity, conversational context, dialogue length, and semantic content. Please see more details in Appendices A-B in our paper.

Citation

If you use this dataset, please cite:

@article{zeng2026bringing,
  title={Bringing Up a Bilingual BabyLM: Investigating Multilingual Language Acquisition Using Small-Scale Models},
  author={Zeng, Linda and Feng, Steven Y and Frank, Michael C},
  journal={arXiv preprint arXiv:2603.29552},
  year={2026}
}

Dataset Card Authors

Linda Zeng, Stanford University

Dataset Card Contact

lindazeng@stanford.edu