From cf7db78486a2445326b3fd07c83661d85fd56969 Mon Sep 17 00:00:00 2001 From: Ronald Tse Date: Tue, 6 Oct 2026 01:10:57 +0800 Subject: [PATCH] docs: curated card for ara-diac-plane-large-1.0 --- cards/ara-diac-plane-large-1.0.md | 36 +++++++++++++++++++++++++++++++ 1 file changed, 36 insertions(+) create mode 100644 cards/ara-diac-plane-large-1.0.md diff --git a/cards/ara-diac-plane-large-1.0.md b/cards/ara-diac-plane-large-1.0.md new file mode 100644 index 0000000..8cd2a62 --- /dev/null +++ b/cards/ara-diac-plane-large-1.0.md @@ -0,0 +1,36 @@ +## Benchmark + +Full-set SadeedDiac-25 — 1,200 paragraphs, zero skipped, greedy decode +(protocol-matched, Misraj's evaluator): + +| model | Total DER | notes | +|---|---|---| +| ara-diac-2.0 (teacher, 580M seq2seq) | 2.29% | best dedicated | +| **ara-diac-plane-large-1.0 (this model)** | **2.74%** | **866 MB int8** | +| Gemini-Flash-2.0 (frontier LLM) | 3.19% | | +| ara-diac-plane-1.0 (plane small) | 3.59% | 219 MB int8 | +| ara-diac-small-2.1 (seq2seq student) | 4.57% | 491 MB int8 | + +The second-best dedicated system overall, ahead of Gemini-Flash, and +the best that ships as one self-contained int8 ONNX graph. Morph DER +1.67% (iʿrāb quality); total WER 8.61%. + +## Architecture + +Plane-factorized diacritization at the ByT5-LARGE encoder scale: a +byte-level encoder-only transformer (no decoder) with a diacritic-plane +embedding and one classification head over 54 haraqat classes per +character position. Decoding is K=2 mask-predict passes — parallel +forwards feeding the predicted plane back; no KV cache, no +autoregressive loop. Skeleton round-trip is byte-exact. + +The artifact is a single self-contained int8 ONNX graph in the plane +zip contract (metadata.yaml with member sha256s + plane.onnx + +classes.json), loadable by secryst-py `PlaneModel.from_zip`. The +canonical distribution is the Secryst IMF artifact (sha256-verified); +this repository mirrors those exact bytes. + +Provenance: rababa run-019, 2 epochs over 551k unique units. The +epoch-2 number is the corrected measurement — the first epoch-2 eval +silently re-scored epoch-1 predictions and was re-run in full (see the +interscript-ml RESULTS ledger).