Skip to content

ara-diac-layerdrop-1.0

Choose a tag to compare

@ronaldtse ronaldtse released this 04 Sep 08:16
· 77 commits to main since this release
df9225a

ara-diac-layerdrop-1.0-int4

IMF v1 (int4, decoder kv, opset 14). Trained from the depth-cut rung: ByT5-small with the encoder halved 12->6 (surviving layers copied VERBATIM from pretraining - both width-cut approaches collapsed at 74.68/82.96 while this works), Muon optimizer, r7 teacher labels (the 2.0 recipe), 6 epochs. ~190M parameters (63% of ByT5-small). Full-set windowed DER-CE 5.784, delta CI [3.033, 3.491] vs in-run teacher 2.2921 — the depth premium over the full-depth G2a peer (4.5701, CI [1.911, 2.352]) is 1.21pp with non-overlapping intervals. The int4 variant (~95MB) is the browser-budget tier; margin-gated before ship. Checkpoint rababa-checkpoints:/rababa_arabic_distill_small/run-009-layerdrop-6ep/best..

field value
task diacritization (Arab → Arab)
artifact ara-diac-layerdrop-1.0-int4.zip (0.14 GiB)
der_teacher_fullset 2.2921 — windowed DER-CE (1400-byte windows, word-boundary split, greedy, haraqat-projected, Misraj evaluator); full 1,200-paragraph SadeedDiac-25
der_student_fullset 5.784 — same harness; paired bootstrap delta 3.2455 [3.033, 3.491]
parity cer_delta 0.0589pp on 2480 samples
sha256 352e02105645a6e283e2c3602505bea0e90085b741666fd863e478e4835e7bc9
license BSD-3-Clause

Runtimes reassemble split parts transparently and verify every sha256:

from secryst import Model
model = Model.load("ara-diac-layerdrop-1.0-int4")