diff --git a/requirements.txt b/requirements.txt
index 72dfd2e9..4e272f33 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -14,7 +14,7 @@ pyyaml>=6
 openai>=1.0
 anthropic>=0.40
 chromadb>=1.0.0
-datasets>=2.16,<3
+datasets>=3
 sentence-transformers>=2.7
 sqlmodel>=0.0.24
 arize-phoenix>=4.0
diff --git a/src/eval/datasets/squad_v2.py b/src/eval/datasets/squad_v2.py
index 360ae70b..f2dd5061 100644
--- a/src/eval/datasets/squad_v2.py
+++ b/src/eval/datasets/squad_v2.py
@@ -63,7 +63,10 @@ def sample_and_freeze(
     from datasets import load_dataset
 
     logger.info("Loading squad_v2 validation split (caches in ~/.cache/huggingface)...")
-    ds = load_dataset("squad_v2", split="validation")
+    # WHY the namespaced id: huggingface_hub 1.x rejects bare dataset ids
+    # ("squad_v2"), requiring the canonical "namespace/name" form. Paired with
+    # datasets>=3, this restores loading on current dependency versions (#18).
+    ds = load_dataset("rajpurkar/squad_v2", split="validation")
     shuffled = ds.shuffle(seed=seed)
     sample = shuffled.select(range(sample_size))