diff --git a/CHANGELOG.md b/CHANGELOG.md
index a689d722..7f65707c 100644
--- a/CHANGELOG.md
+++ b/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)
diff --git a/pyproject.toml b/pyproject.toml
index 46b66d77..97998892 100644
--- a/pyproject.toml
+++ b/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",
@@ -49,6 +50,7 @@ dependencies = [
     "websockets>=15.0.1",
     "fastapi>=0.116.1",
     "pytest>=6.0.0",
+    "pytest-asyncio>=0.21.0",
     "peewee>=3.18.2",
 ]
 
@@ -59,7 +61,6 @@ Homepage = "https://github.com/fireworks-ai/eval-protocol"
 dev = [
     "build",
     "twine",
-    "pytest-asyncio",
     "pytest-httpserver",
     "werkzeug>=2.0.0",
     "ruff>=0.5.0",
@@ -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 = [
diff --git a/uv.lock b/uv.lock
index 6758d24c..035b9f37 100644
--- a/uv.lock
+++ b/uv.lock
@@ -1108,6 +1108,7 @@ dependencies = [
     { name = "psutil" },
     { name = "pydantic" },
     { name = "pytest" },
+    { name = "pytest-asyncio" },
     { name = "python-dotenv" },
     { name = "pyyaml" },
     { name = "requests" },
@@ -1143,7 +1144,6 @@ dev = [
     { name = "pip" },
     { name = "pre-commit" },
     { name = "pyright" },
-    { name = "pytest-asyncio" },
     { name = "pytest-cov" },
     { name = "pytest-httpserver" },
     { name = "pytest-xdist" },
@@ -1198,9 +1198,9 @@ requires-dist = [
     { name = "anthropic", specifier = ">=0.59.0" },
     { name = "build", marker = "extra == 'dev'" },
     { name = "dataclasses-json", specifier = ">=0.5.7" },
-    { name = "datasets" },
-    { name = "datasets", marker = "extra == 'adapters'", specifier = ">=2.0.0" },
-    { name = "datasets", marker = "extra == 'huggingface'", specifier = ">=2.0.0" },
+    { name = "datasets", specifier = ">=3.0.0" },
+    { name = "datasets", marker = "extra == 'adapters'", specifier = ">=3.0.0" },
+    { name = "datasets", marker = "extra == 'huggingface'", specifier = ">=3.0.0" },
     { name = "deepdiff", specifier = ">=6.0.0" },
     { name = "docker", marker = "extra == 'dev'", specifier = "==7.1.0" },
     { name = "docstring-parser", specifier = ">=0.15" },
@@ -1238,7 +1238,7 @@ requires-dist = [
     { name = "pydantic", specifier = ">=2.0.0" },
     { name = "pyright", marker = "extra == 'dev'", specifier = ">=1.1.365" },
     { name = "pytest", specifier = ">=6.0.0" },
-    { name = "pytest-asyncio", marker = "extra == 'dev'" },
+    { name = "pytest-asyncio", specifier = ">=0.21.0" },
     { name = "pytest-cov", marker = "extra == 'dev'" },
     { name = "pytest-httpserver", marker = "extra == 'dev'" },
     { name = "pytest-xdist", marker = "extra == 'dev'" },