Fused-trunk multitask model with a low-rank per-setup head - #114

Merged

RobbinBouwmeester merged 9 commits into

mainfrom

feat/multitask-flexcnn-v7b

Aug 20, 2026

Commits on Aug 20, 2026

feat: fused-trunk multitask model with a low-rank per-setup head

RobbinBouwmeester
and
claude
committed
2c7ba60
style: apply ruff format

RobbinBouwmeester
and
claude
committed
c74e354
fix: address review on memory, serialisation, device and feature spec

RobbinBouwmeester
and
claude
committed
5b14ad3
feat: adapt a fused-trunk model to a new setup by fitting its low-rank head

RobbinBouwmeester
and
claude
committed
3a2e08d
fix: keep fine-tuning from collapsing on a small reference set

RobbinBouwmeester
and
claude
committed
fb5ff0e
fix: stop fine-tuning returning a model worse than its starting point

RobbinBouwmeester
and
claude
committed
98ac37c
docs(data): untangle the interleaved parameter descriptions

RobbinBouwmeester
and
claude
committed
2de8825
feat(features): add legacy_positional_deltas for models trained before 4.0.1

RobbinBouwmeester
and
claude
committed
28eb892
fix(features): keep pre-4.0.1 models on the encoding they were trained on

RobbinBouwmeester
and
claude
committed
ab77e6d