The BabyLM Challenge asks models to learn from developmentally plausible data, but how closely the standard corpus resembles the language infants actually hear has not been measured. We compare the Strict-Small corpus against SEEDLingS, a corpus of day-long home recordings from 44 infants with 358,300 hand-annotated noun tokens, over the 299 words the two share. Corpus word frequencies match real infant input only moderately (Spearman’s 𝜌 = 0.42; all correlations reported here are Spearman rank correlations), and real input frequency predicts human age of acquisition (AoA) far better (𝜌 = -0.79) than the training corpus’s own frequencies do (𝜌 = -0.48).

Breaking the corpus down by source locates the gap. CHILDES tracks real infant input at 𝜌 = 0.55 while the five other sources fall between 0.14 and 0.31, and the pooled corpus falls below CHILDES alone. The sources are not ordered by how speech-like they are: written text ranks above adult conversation. What separates them is child-directedness.

Against this we set a curriculum intervention. We order the corpus by the AoA of the words each sentence contains, derived from the MacArthur-Bates Communicative Development Inventory (CDI), train GPT-2 Small on it and on a content-matched shuffled baseline, across three seeds and three training budgets up to 100M words. The curriculum reliably produces a model whose acquisition order matches the human one (𝜌 = 0.63–0.78 against 0.09–0.22), but the curriculum encodes human AoA at 𝜌 = 0.96 by construction and the alignment falls to 𝜌 ≈ 0.3 once presentation order is partialled out. It confers no advantage on BLiMP at any budget. Improving what a corpus contains may matter more than the order in which it is presented.