ILM and ArfaLM: Child-Scale Language Models Trained on Natural Parent-Child Speech for Deployment on Sub-1GHz Legacy Hardware. Part 1: Models, Data, and Architectural Comparison

We present ILM and ArfaLM, two language models trained on naturalistic child-directed speech for deployment on legacy 1990s computing hardware. ILM is a 2.3 million parameter, two-layer LSTM that uses only architectural and training techniques available before the year 2000. ArfaLM is an 8.4 million parameter, six-layer decoder-only Transformer trained with modern methods on the same corpus. Both models are trained on 250,757 filtered conversation pairs extracted from 2,930 sessions in the CHILDES Eng-UK corpus, with 5 complete child histories (93 sessions, 14,504 pairs) held out for true generalization testing. We also evaluate against a 3-gram smoothed baseline trained on the same data. On the held-out test set of 5 children that no model has seen during training, ArfaLM achieves vocabulary-normalized per-word perplexity 26.39 and ILM-SGD achieves 56.62, compared to 417.21 for the 3-gram baseline. ArfaLM is therefore 15.8 times lower-perplexity than the n-gram baseline and 2.15 times lower-perplexity than the era-authentic LSTM on identical data. An ablation that uses ILM's LSTM with the modern AdamW optimiser (ILM-AdamW, test perplexity 36.33) decomposes the 2.15 times gain into 1.56 times from the optimiser change alone and 1.38 times from the bundled effects of architecture, tokenization, and parameter scale. Notably, the optimizer change contributes more to the total gain than the architectural change does on this corpus at this scale. Generalization gaps are 1.12 for ArfaLM, 0.91 for ILM-AdamW (the best of any model, indicating the flattest minimum), and 1.01 for ILM-SGD, indicating strong within-corpus generalization to unseen children with no evidence of overfitting at the 2-tokens-per-parameter ratio. A Gutenberg literature ablation produces a clean null result: adding 19th-century children's literature to CHILDES neither helps nor hurts measurably. ILM is dedicated to the memory of Imran Shah (1969 to 2019). ArfaLM is named for Arfa Karim Randhawa (1995 to 2012), the youngest Microsoft Certified Professional. This is Part 1 of two. Part 2 will report C inference engine optimization, external safety filtering, and deployment on original IBM 300PL hardware. Artifacts:- Dataset: https://huggingface.co/datasets/nshah-fbcs/childes-engUK-conversational-pairs- Models: https://huggingface.co/nshah-fbcs/ilm-arfalm-childes