Pages, glossary terms, models, and whatever the page in front of you is showing.
The bigger sibling of the 13.77M model next door. Use it to compare two Bulgarian sentences, continue a prompt, or trace one short continuation. The model is downloaded once and runs in your tab; your text does not leave your machine. Loading the model…
This is the model whose 92.3% grammar-preference number the smaller page cites — measured here directly, not borrowed. Give it two sentences and it reports how surprised it is by each (, lower = more natural); on 26 real Bulgarian error/correction pairs, it picks the correct sentence 92.3% of the time.
More capacity buys more locally fluent Bulgarian, not more honesty: this is still a base model with no instruction tuning, so it continues text rather than answering it, and it still invents facts with total confidence. It is here so the judge number above has something to contrast against, not because generation is the point — the 13.77M page makes the identical claim at 6× less download.
The same viewer as the smaller model and the restorer — attention arcs, a , and attribution — now pointed at a deeper model. A third, ~124 MB file loads the first time you trace something.
Same question the smaller page asks, answered again at this size rather than assumed to generalize: does shrinking the model file make the model worse?
Measured on the same external Bulgarian held-out every number in this project uses: the uncompressed graph scores 1.2813 bits/char; the compressed graph, 1.2849 — a cost of +0.0036 bpc, larger than the 13.77M model's +0.0018 but still small enough that the extra training stage was unnecessary.
Per-position agreement between the two graphs is 97.1% (35 probed positions) — not the near-100% the smaller model showed, which is worth stating plainly rather than rounding away: compression costs a little more at this width, not nothing.
