Here's what I found. Yes, there's meaningful work in this area — though the literature is fragmented and hasn't consolidated into a single well-known "extraction pipeline" yet. Here are the relevant findings:
This is the most directly relevant line of work. The BabyLM Corpus is a ~100M-word, developmentally-driven dataset that mirrors children's linguistic input (child-directed speech, caregiver dialogue transcripts, children's literature). The BabyLM Challenge (started ACL 2024, with a 3rd edition in 2025) benchmarks sample-efficient models trained on cognitively plausible data.
Key papers:
- "Sample-efficient pretraining on a developmentally plausible corpus" (BabyLM, ACL 2024)
- "Findings of the third BabyLM challenge: accelerating language modeling research with cognitively plausible data" (2025)