Induction heads in a browser tab: a hand-written WebGPU trainer, and what the repeat rule and weight tying do to onset
A small transformer trains through the induction transition in a browser tab, on the reader's GPU, and reaches the same event step as PyTorch on 64 of 64 seeds. A seeds study on small models asks what sets the onset of induction and whether tying weights delays it. Code, logs and the command behind every number are in the repository.
Authors
- Abhay Mettu (ORCID: https://orcid.org/0009-0001-2434-353X)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-02
- DOI
- https://doi.org/10.5281/zenodo.23899099
- Primary Topic
- Topic Modeling
- Type
- preprint