update dataset version requirement

Author: benjibc

CHANGELOG.md

@@ -20,7 +20,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - (Your now removed features here)
 
 ### Fixed
-- (Your bug fixes here)
+- Pin minimum `datasets` version to `>=3.0.0` to avoid import failures with `pyarrow>=21` (AttributeError: `pyarrow.PyExtensionType`). This ensures compatibility when installing via pip/uv in clean CI environments.
 
 ### Security
 - (Your vulnerabilities patched here)

pyproject.toml

@@ -27,7 +27,8 @@ dependencies = [
     "aiohttp",
     "mcp>=1.9.2",
     "PyYAML>=5.0",
-    "datasets",
+    # Pin minimum datasets to avoid pyarrow API mismatch (PyExtensionType removal in pyarrow>=21)
+    "datasets>=3.0.0",
     "fsspec",
     "hydra-core>=1.3.2",
     "omegaconf>=2.3.0",
@@ -103,12 +104,14 @@ langfuse = [
     "langfuse>=2.0.0",
 ]
 huggingface = [
-    "datasets>=2.0.0",
+    # Keep in sync with core dependency to ensure compatibility with latest pyarrow
+    "datasets>=3.0.0",
     "transformers>=4.0.0",
 ]
 adapters = [
     "langfuse>=2.0.0",
-    "datasets>=2.0.0",
+    # Keep in sync with core dependency to ensure compatibility with latest pyarrow
+    "datasets>=3.0.0",
     "transformers>=4.0.0",
 ]
 bigquery = [