bbunzeck/generation-evals-training-data

This repository contains the pretraining data for the models probed in Child-directed speech facilitates production, not comprehension, in BabyLMs (Bunzeck & Zarrieß @ CoNLL 2026).

If you use this data, please cite:

@inproceedings{bunzeck-zarriess-2026-child,
    title = "Child-directed speech facilitates production, not comprehension, in {B}aby{LM}s",
    author = "Bunzeck, Bastian  and
      Zarrie{\ss}, Sina",
    editor = "Bonial, Claire  and
      Berzak, Yevgeni",
    booktitle = "Proceedings of the 30th Conference on Computational Natural Language Learning",
    month = jul,
    year = "2026",
    address = "San Diego, California, USA",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.conll-main.14/",
    doi = "10.18653/v1/2026.conll-main.14",
    pages = "227--249",
    ISBN = "979-8-89176-410-1",
    abstract = "Recent studies suggest that child-directed speech is not conducive to language learning in BabyLMs. However, current evaluations focus predominantly on comprehension and not production, which is central to usage-based theories of language acquisition which argue how CDS facilitates early language use through constructional ``frames'' (frequent lexical patterns with open slots). We introduce a novel generation-based evaluation inspired by such theories in form of a \textbf{frame-completion task}, and compare Llama models trained with CDS, the BabyLM corpus, and web-crawl data (FineWeb-edu) on comprehension benchmarks and our novel framework. Our results reveal a clear dissociation between models' comprehension and production capabilities: while FineWeb-trained models excel at minimal pairs, CDS-trained models produce grammatical completions substantially earlier in training and concentrate probability mass on appropriate slot-fillers. These findings show that comprehension benchmarks underestimate what CDS affords to BabyLMs."
}