chore: Migrate to transformers v5 (with v4 fallback for distill)

.github/workflows/test.yml

@@ -49,7 +49,10 @@ jobs:
         source .venv/bin/activate
         uv pip install ./dist/sieves-*.whl
         # Optional dependencies need to be installed in separate command.
-        uv pip install 'sieves[ingestion,distill,test]'
+        # Note: due to `setfit` not supporting `transformers` v5 yet, installing the `distill` extra runs with v4 of
+        # `transformers`. As we're installing all packages for tests, we're running the test suite with `transformers`
+        # v4. This is an ok tradeoff for now, given the minimal integration surface of `transformers` within `sieves`.
+        uv pip install 'sieves[ingestion,distill,optimization,test]'
 
     - name: Build docs
       run: |

CITATION.CFF

@@ -15,7 +15,7 @@ authors:
 - family-names: "Maufe"
   given-names: "Matthew"
 title: "Sieves"
-version: 1.0.2
+version: 1.0.3
 doi: 10.5281/zenodo.18664061
-date-released: 2026-02-16
+date-released: 2026-05-01
 url: "https://github.com/MantisAI/sieves"

README.md

@@ -236,8 +236,9 @@ pip install sieves
 
 **Optional extras:**
 ```bash
-pip install "sieves[ingestion]"  # PDF/DOCX parsing (docling, marker)
-pip install "sieves[distill]"    # Model distillation (setfit, model2vec)
+pip install "sieves[ingestion]"    # PDF/DOCX parsing (docling, marker)
+pip install "sieves[distill]"      # Model distillation (setfit, model2vec)
+pip install "sieves[optimization]" # Prompt optimization with DSPy (optuna)
 ```
 
 ## Community & Support

pyproject.toml

@@ -26,7 +26,7 @@ dependencies = [
     # Model wrappers
     "outlines>=1,<2",
     "gliner2>=1.2,<2",
-    "transformers[torch]>=4,<5",
+    "transformers[torch]>=4,<6",
     "dspy-ai>=3,<4",
     "dspy>=3,<4",
     "accelerate>1.2,<2",
@@ -54,13 +54,20 @@ Homepage = "https://github.com/MantisAI/sieves"
 Repository = "https://github.com/MantisAI/sieves"
 
 [project.optional-dependencies]
+optimization = [
+    "optuna>=4,<5",
+]
 ingestion = [
     "docling>=2,<3",
     "nltk>=3.9.1",
 ]
 distill = [
     "setfit>=1.1,<2",
     "model2vec[train]>0.4,<0.5",
+    # setfit 1.1.3 (latest, Nov 2024) imports `default_logdir` from
+    # transformers.training_args, which was removed in transformers v5.
+    # Cap distill to transformers v4 until setfit ships v5 support.
+    "transformers[torch]>=4,<5",
 ]
 test = [
     "anthropic>=0.45,<1",
@@ -130,6 +137,9 @@ ignore = ["D203", "D212"]
 fixable = ["ALL"]
 unfixable = ["F401"]
 
+[tool.uv]
+exclude-newer = "3 days"
+
 [tool.ruff.lint.pydocstyle]
 convention = "pep257"
 

sieves/model_wrappers/huggingface_.py

@@ -66,7 +66,7 @@ def execute(values: Sequence[dict[str, Any]]) -> Sequence[tuple[Result | None, A
                         sequences=[doc_values["text"] for doc_values in values],
                         candidate_labels=prompt_signature,
                         hypothesis_template=template,
-                        mode="multi",
+                        multi_label=True,
                         **self._inference_kwargs,
                     )
 

sieves/serialization.py

@@ -79,7 +79,7 @@ def get_version() -> str:
 
         :return str: Version string from setup.cfg metadata.
         """
-        return "1.0.2"
+        return "1.0.3"
 
     version: str = get_version()
     cls_name: str