EML-LM — large variants
Browser-loadable checkpoints for the EML Foundation lab app on
The Daily Synthesis.
Trained by eml-autoresearch/ and exported with eml-autoresearch/export.py;
scripts/eml-foundation/eml-lm-store.js imports them straight into IndexedDB.
| file | size | arch | precision | val_bpb | why |
|---|---|---|---|---|---|
eml-lm-default.json |
59 MB | depth 4 | f32 | 2.348 | byte-exact MLX parity for researchers |
eml-lm-depth6-int8.json |
35 MB | depth 6 | int8 | 2.564 | alternative per-layer k profile |
The default download is not here. eml-lm-depth4-int8.json (15 MB, depth 4,
int8, val_bpb 2.348) is argmax-lossless against MLX at a quarter of the size, so
it is the first entry in the app's AUTO_IMPORT_SOURCES and the one nearly every
visitor gets. It ships from the site itself.
These two live here because the published site is bound by GitHub Pages' 1 GB
limit and these 93 MB were 85% of its models/ directory while being optional
research variants that the auto-import path never reaches.
Regenerate either with eml-autoresearch/export.py --out <name>.json.