Added aviary-paper-data as LitQA v2 test split

.github/workflows/tests.yml

@@ -49,6 +49,10 @@ jobs:
           enable-cache: true
       - run: uv python pin ${{ matrix.python-version }}
       - run: uv sync --python-preference=only-managed
+      - name: Login to Hugging Face Hub
+        run: uv run --no-project huggingface-cli login --token $HUGGINGFACE_HUB_ACCESS_TOKEN
+        env:
+          HUGGINGFACE_HUB_ACCESS_TOKEN: ${{ secrets.HUGGINGFACE_HUB_ACCESS_TOKEN }}
       - name: Cache datasets
         uses: actions/cache@v4
         with:

paperqa/agents/task.py

@@ -39,6 +39,7 @@
 )
 from paperqa.docs import Docs
 from paperqa.litqa import (
+    DEFAULT_AVIARY_PAPER_HF_HUB_NAME,
     DEFAULT_LABBENCH_HF_HUB_NAME,
     DEFAULT_REWARD_MAPPING,
     read_litqa_v2_from_hub,
@@ -408,6 +409,7 @@ def compute_trajectory_metrics(
 class LitQAv2TaskSplit(StrEnum):
     TRAIN = "train"
     EVAL = "eval"
+    TEST = "test"
 
 
 class LitQAv2TaskDataset(LitQATaskDataset):
@@ -416,20 +418,23 @@ class LitQAv2TaskDataset(LitQATaskDataset):
     def __init__(
         self,
         *args,
-        labbench_dataset: str = DEFAULT_LABBENCH_HF_HUB_NAME,
+        train_eval_dataset: str = DEFAULT_LABBENCH_HF_HUB_NAME,
+        test_dataset: str = DEFAULT_AVIARY_PAPER_HF_HUB_NAME,
         read_data_kwargs: Mapping[str, Any] | None = None,
         split: str | LitQAv2TaskSplit = LitQAv2TaskSplit.EVAL,
         **kwargs,
     ):
         super().__init__(*args, **kwargs)
-        train_df, eval_df = read_litqa_v2_from_hub(
-            labbench_dataset, **(read_data_kwargs or {})
+        train_df, eval_df, test_df = read_litqa_v2_from_hub(
+            train_eval_dataset, test_dataset, **(read_data_kwargs or {})
         )
         split = LitQAv2TaskSplit(split)
         if split == LitQAv2TaskSplit.TRAIN:
             self.data = train_df
         elif split == LitQAv2TaskSplit.EVAL:
             self.data = eval_df
+        elif split == LitQAv2TaskSplit.TEST:
+            self.data = test_df
         else:
             assert_never(split)
 

paperqa/litqa.py

@@ -37,26 +37,30 @@ def make_discounted_returns(
 
 
 DEFAULT_LABBENCH_HF_HUB_NAME = "futurehouse/lab-bench"
+# Test split from Aviary paper's section 4.3: https://doi.org/10.48550/arXiv.2412.21154
+DEFAULT_AVIARY_PAPER_HF_HUB_NAME = "futurehouse/aviary-paper-data"
 
 
 def read_litqa_v2_from_hub(
-    labbench_dataset: str = DEFAULT_LABBENCH_HF_HUB_NAME,
+    train_eval_dataset: str = DEFAULT_LABBENCH_HF_HUB_NAME,
+    test_dataset: str = DEFAULT_AVIARY_PAPER_HF_HUB_NAME,
     randomize: bool = True,
     seed: int | None = None,
     train_eval_split: float = 0.8,
-) -> tuple[pd.DataFrame, pd.DataFrame]:
+) -> tuple[pd.DataFrame, pd.DataFrame, pd.DataFrame]:
     """
-    Read LitQA v2 JSONL into train and eval DataFrames.
+    Read LitQA v2 JSONL into train, eval, and test DataFrames.
 
     Args:
-        labbench_dataset: The Hugging Face Hub dataset's name corresponding with the
-            LAB-Bench dataset.
+        train_eval_dataset: Hugging Face Hub dataset's name corresponding with train
+            and eval splits.
+        test_dataset: Hugging Face Hub dataset's name corresponding with a test split.
         randomize: Opt-out flag to shuffle the dataset after loading in by question.
         seed: Random seed to use for the shuffling.
         train_eval_split: Train/eval split fraction, default is 80% train 20% eval.
 
     Raises:
-        DatasetNotFoundError: If the LAB-Bench dataset is not found, or the
+        DatasetNotFoundError: If any of the datasets are not found, or the
             user is unauthenticated.
     """
     try:
@@ -67,9 +71,15 @@ def read_litqa_v2_from_hub(
             " `pip install paper-qa[datasets]`."
         ) from exc
 
-    litqa_v2 = load_dataset(labbench_dataset, "LitQA2")["train"].to_pandas()
-    litqa_v2["distractors"] = litqa_v2["distractors"].apply(list)
+    train_eval = load_dataset(train_eval_dataset, "LitQA2")["train"].to_pandas()
+    test = load_dataset(test_dataset, "LitQA2")["test"].to_pandas()
+    # Convert to list so it's not unexpectedly a numpy array
+    train_eval["distractors"] = train_eval["distractors"].apply(list)
+    test["distractors"] = test["distractors"].apply(list)
+    # Let downstream usage in the TaskDataset's environment factories check for the
+    # presence of other DataFrame columns
     if randomize:
-        litqa_v2 = litqa_v2.sample(frac=1, random_state=seed)
-    num_train = int(len(litqa_v2) * train_eval_split)
-    return litqa_v2[:num_train], litqa_v2[num_train:]
+        train_eval = train_eval.sample(frac=1, random_state=seed)
+        test = test.sample(frac=1, random_state=seed)
+    num_train = int(len(train_eval) * train_eval_split)
+    return train_eval[:num_train], train_eval[num_train:], test

tests/test_litqa.py

@@ -25,7 +25,7 @@ def test_make_discounted_returns(
 
 def test_creating_litqa_questions() -> None:
     """Test making LitQA eval questions after downloading from Hugging Face Hub."""
-    _, eval_split = read_litqa_v2_from_hub(seed=42)
+    eval_split = read_litqa_v2_from_hub(seed=42)[1]
     assert len(eval_split) > 3
     assert [
         MultipleChoiceQuestion(

tests/test_task.py

@@ -98,7 +98,11 @@ class TestTaskDataset:
 
     @pytest.mark.parametrize(
         ("split", "expected_length"),
-        [(LitQAv2TaskSplit.TRAIN, 159), (LitQAv2TaskSplit.EVAL, 40)],
+        [
+            (LitQAv2TaskSplit.TRAIN, 159),
+            (LitQAv2TaskSplit.EVAL, 40),
+            (LitQAv2TaskSplit.TEST, 49),
+        ],
     )
     @pytest.mark.asyncio
     async def test___len__(