StemmaLM weights

Snapshots of the StemmaLM individuals (site: https://stemmalm.github.io): small character-level language models raised from scratch on adult speech from AO-CHILDES, in order of the child's age, one snapshot per month bin (m00 is before any learning). Layout: n001/vocab.json, n001/meta.json, n001/ckpt/mXX.pt (fp16 PyTorch state dicts), and n001/web/ (the same snapshots as flat fp16 files plus manifest.json, read by the in-browser model on the site).

The models can reproduce names from the CHILDES transcripts. The site replaces names in replies with “(name)”; generating directly from these weights does not.

Training data: CHILDES (MacWhinney 2000) via AO-CHILDES (Huebner & Willits 2021), CC BY-NC-SA 4.0. Non-commercial use only.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support