Model zoo

Models you can scry with

Ids resolve through the models.yaml index. Teachers are the strongest trained models; students are the compact distillations that actually ship in the crystals.

Zoo

id / familyrevealsmeasured quality
ara-diac-1.0teacherArabic consonants → harakāt2.58% windowed DER-CE on SadeedDiac-25; 1.53% on word-final iʿrāb endings
ara-diac-small-1.0studentArabic → harakāt (compact, 1.3 GiB)3.66% on the same harness — the +2.34pp cost of the ByT5-small client rung
heb-diac-1.1studentHebrew consonants → nikud16.44% DER greedy / 16.43% beam-4 — the current best Hebrew rung
heb-diac-1.0studentHebrew → nikud17.46% DER beam-4 (DictaBERT baseline 35.6%)
heb-diac-small-1.0studentHebrew → nikud (compact)30.37% DER vs teacher 24.79% — +5.6pp for the 2× shrink
fas-g2p-1.0studentPersian → Latin reading1.6% CER; 77.34% SentenceBench homograph accuracy
tha-g2p-base-1.0studentThai → IPAteacher at 4.43% PER (public baseline 6.37%); student gated at +4.76pp
tha-g2p-small-1.0studentThai → IPA (int8, 246 MiB)2.85% PER greedy on the runtime protocol; parity 0.08pp
urd-diac-1.0studentUrdu → vocalized text3.74% CER
urd-g2p-1.0studentUrdu → phonemes (IPA)14.77% CER — first learned Urdu baselines we know of
khm-latn-1.0studentKhmer text → Latin reading59.66% exact match, 27.42 CER (forward model)

11 of 11 entries

Where the numbers come from

Every metric above was produced by the benchmark’s own evaluator on its full test set — Misraj’s evaluator for SadeedDiac-25, SentenceBench for Persian homographs, the fixed 1,219-sentence Kaikki set for Thai — under published, zero-skip protocols. Methodology and run IDs live with each project’s results file; nothing here is self-graded on a private split.

Conformance methodology

Teachers are trained in secryst-train (secryst’s own monorepo); distilled students that pass the export gate are published into the interscript-ml zoo. Before any model enters the index, the gate writes its parity record into the artifact: the student is diffed against its teacher on held-out samples, and a model whose behavior differs from its published identity does not ship. At the runtime level, the three crystals are diffed against each other on shared golden sets (generated by the Python crystal) in CI — the release gate for gem 1.0 and every subsequent crystal release.