Skip to content

Commit b73892d

Browse files
author
Ronald Tse
committed
docs: ara-diac-small-2.1 README + registry pointer
The 2.1 rung's release notes (4.5701 [1.91,2.35], epochs lever, frontier context) and the export registry entry now points at its own README instead of borrowing 2.0's.
1 parent 9fc02f6 commit b73892d

2 files changed

Lines changed: 30 additions & 1 deletion

File tree

Lines changed: 29 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,29 @@
1+
# ara-diac-small-2.1
2+
3+
Arabic diacritization (haraqat restoration). Client-tier ByT5-small
4+
student — the 2.0 recipe (r7 canonical teacher, Muon) held for
5+
**double the epochs** (6 instead of 3, 39,018 steps): the epochs lever
6+
moves the rung 4.82 → **4.5701** full-set windowed DER-CE.
7+
8+
The paired-bootstrap delta vs the teacher (same harness, seed 42,
9+
1,000 resamples) is [1.91, 2.35] — disjoint from the 2.0 rung's
10+
[2.36, 2.82]: the improvement is statistically real, and the frontier
11+
separations hold end to end (1.0 8.26 → lite 5.78 → 2.0 4.82 → 2.1
12+
4.57 → teacher 2.29).
13+
14+
The epochs result reframes the residual: doubling training buys
15+
−0.25pp, so undertraining is not the dominant term — the E2/E3 domain
16+
attribution substantially stands. The strict teacher+0.5pp gate
17+
remains missed (disclosed); the causal follow-ups (E6 register swap —
18+
negative; G2b 5× classical corpus — the add-direction test) are in
19+
docs/RESULTS.md.
20+
21+
```python
22+
from interscript_ml import Model
23+
model = Model.load("ara-diac-small-2.1")
24+
model.translate("قوله فحكمها في الوفاة")
25+
```
26+
27+
Same IMF v1 contract as 2.0: dynamic fetch, sha256-verified, KV
28+
decode, margin report alongside the parity block. Checkpoint:
29+
`rababa_arabic_distill_small/run-007-r7-muon-6ep/best`.

src/gpu/modal_export.py

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -127,7 +127,7 @@
127127
"volume": "/volumes/rababa-checkpoints",
128128
"checkpoint": "rababa_arabic_distill_small/run-007-r7-muon-6ep/best",
129129
"metadata": "models/ara-diac-small/ara-diac-small-2.1.metadata.yaml",
130-
"readme": "models/ara-diac-small/ara-diac-small-2.0.README.md",
130+
"readme": "models/ara-diac-small/ara-diac-small-2.1.README.md",
131131
"test_volume": "/datasets/rababa",
132132
"test_data": "arabic-sadeed-imf/test.jsonl",
133133
"probe": "قوله",

0 commit comments

Comments
 (0)