Guard pretrained transfer, normalise content scale, and fit the standardiser on the full split - #48

Open

Rian354 wants to merge 10 commits into

mainfrom

feat/ssl-pretraining-and-encoders

Commits on Aug 11, 2026

Guard pretrained transfer, normalise content scale, fit the standardiser on the full split
Rian354
committed
c871cdb

Commits on Aug 12, 2026

Mask batch padding in the unified sequence, and read observation flags from the right field
Rian354
committed
6e9a073
Drive the standardiser fit from the region of interest, not the parent index map
Rian354
committed
12add17
Record padding in the collate the dataloader actually uses
Rian354
committed
d770167
Pad every ragged dimension in the collate, not only dimension 0
Rian354
committed
bbfbcf3
Separate the token mask from the padding mask in the unified embedding
Rian354
committed
17aaeb0

Commits on Aug 13, 2026

Key the frozen text cache on real tokens, not on batch padding
Rian354
committed
fb16cc1
Name the standardiser provenance as region_of_interest, not the parent map

Rian354
and
cursoragent
committed
d08b8d7

Commits on Aug 18, 2026

Keep padding_idx=0 on nested code embeddings.

Rian354
and
cursoragent
committed
bc00852
Thread collate pad_mask through the unified MLP path.

Rian354
and
cursoragent
committed
996fea7