Skip to content

Commit 90b79b0

Browse files
committed
docs: r5 paragraph-context verdict — 2.6775/1.5965, beats verified GLM-5.2
Windowed zero-skip at 1400B context: our 580M ByT5 now beats the GLM-5.2 reproduction (2.6911/1.7179) on both Total and Morph DER. Paragraph-unit training closed the context gap; GTPO-GRPO chained on r5 best by the supervisor.
1 parent 7b55991 commit 90b79b0

1 file changed

Lines changed: 1 addition & 0 deletions

File tree

docs/RESULTS.md

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -47,6 +47,7 @@ protocol, zero skipped paragraphs.
4747
| rababa_arabic_byt5 r2 (ours) | 580M | 2.9406 | 1.8333 | 8.8373 | 5.0835 |
4848
| rababa_arabic_byt5 r2 (beam 4) | 580M | 2.9478 | 1.8522 | 8.8143 | 5.1190 |
4949
| **rababa_arabic_v2 (ours)** | **~10M** | **3.2495** | **1.8072** | 10.3276 | **5.2953** |
50+
| **rababa r5 paragraph-context, windowed zero-skip** | 580M | **2.6775** | **1.5965** | **8.0919** | **4.3863** |
5051
| rababa r3 + RAFT run-002 (beam 4) | 580M | 2.8515 | 1.7617 | 8.5410 | 4.8859 |
5152
| rababa r3 + RAFT run-002 (beam 1) | 580M | 2.8308 | 1.7638 |||
5253

0 commit comments

Comments
 (0)