sentence-transformers
[`feat`] Introduce cross-modality and multi-modality support; modularize CrossEncoder class
#3554
Open
Open
[`feat`] Introduce cross-modality and multi-modality support; modularize CrossEncoder class #3554
tomaarsen
wants to merge
46 commits into
huggingface:main
from
tomaarsen:refactor/multimodal
Introduce cross-modality and multi-modality support
c0af86c0
Heavily refactor Router to allow for modality-based routing
2273532a
Fix Router docstring examples
efb36f7a
Remove now-removed Transformer types
6ae100e6
MLMTransformer just as Transformer with transformer_task="fill-mask"
b20da09d
Add TODOs to modality_utils
cbc7f743
Merge branch 'main' into refactor/multimodal
0aa54b9f
Rename tokenize to preprocess, soft deprecation
0ede3859
Load with string instead of Path
4c54d2ab
Heavily expand on refactor: separate "Base..." classes to act as supe…
6c61c1d6
Avoid ImportError
a7abc9ee
Work on matching performance of pre-refactor for CrossEncoder
8ba0b3b7
Let's stick with PretrainedConfig instead of PreTrainedConfig for now
0df4450f
Improve CrossEncoderTrainer; remove router_mapping from CE
bff6efac
Improve the Router, its tests now pass
a55f94b0
Modernize the CMNRL test: avoid InputExample/smart_batching_collate
688bbe98
Require v5.0.0 (pre)
9c47fcac
Make max_seq_length more robust, update model type default
805d11e8
Update monkeypatch for hard negatives test
2cdb3429
Lower tokenizer model_max_length if needed; fix CE model card test
8df59559
Let's revert to <5 in the CI, just because optimum isn't compatible yet
9c34b7f5
Move multi-processing functionality to base, fix it for SparseEncoder
aae848ec
Merge branch 'main' into refactor/multimodal
59c39f4b
Use warmup_steps instead of deprecated warmup_ratio
aefa87ae
Merge branch 'main' into refactor/multimodal
64d2b39f
Load with fp32 to match existing test expectation
30623050
Only run tests for transformers v5+, install accelerate from source
1b061109
Improve prompts support, use chat_template with rerankers
7331265c
CE: Load prompt and default_prompt_name from config again
16804f76
CE: Correctly grab logits in CausalScoreHead if padding is left
ab4abaa8
Add some comments to self
11480e7c
copilot-pull-request-reviewer
commented
on 2025-12-15
Add TODO in code
bd5e23b4
copilot-pull-request-reviewer
commented
on 2025-12-16
Merge branch 'main' into refactor/multimodal
14395489
Extend "messages" support for all model archetypes, uses chat_template
a91ab2de
Update the ruff pre-commit hooks slightly
181bda8a
Add show-fixes to get an idea of what's meant to be failing in pre-co…
31c6e35c
Update the Pooling import
b3cb2d86
Formalize deprecated imports
2422a2d7
Add simple __repr__ to Module superclass
a5ae50f3
Revert to simply sys.modules aliasing to avoid different instances o…
2d23def8
Ignore unresolved-attribute with ty
ea61bef1
Work towards transformers v4 compatibility as well
56ed228d
Patch Asym import
a9751216
Fix typo in module output name
0425b442
Also use unreleased accelerate for transormers <5
1cb7d10c
Merge branch 'main' into refactor/multimodal
3c55e8b1
Reviewers
Assignees
No one assigned
Labels
None yet
Milestone
No milestone
Login to write a write a comment.
Login via GitHub