Exposed seeding of LitQA2 read and shuffling

paperqa/agents/task.py

@@ -196,6 +196,7 @@ def __init__(
         base_query: QueryRequest | dict | None = None,
         base_docs: Docs | dict | None = None,
         rewards: Mapping[str, float] = DEFAULT_REWARD_MAPPING,
+        question_kwargs: Mapping[str, Any] | None = None,
         eval_model: LLMModel | str = DEFAULT_EVAL_MODEL_NAME,
         **env_kwargs,
     ):
@@ -210,23 +211,23 @@ def __init__(
             base_docs = Docs(**base_docs)
         self._base_docs = base_docs
         self._rewards = rewards
-        self._env_kwargs = env_kwargs
+        self._question_kwargs = question_kwargs
         self._eval_model = eval_model
+        self._env_kwargs = env_kwargs
 
     def _make_gradable_environment(
         self,
         ideal: str,
         distractors: str | list[str],
         question: str,
-        use_unsure: bool = True,
         sources: str | list[str] | None = None,
     ) -> GradablePaperQAEnvironment:
         qa_prompt, evaluation_from_answer = LitQAEvaluation.from_question(
             ideal=ideal,
             distractors=distractors,
             question=question,
-            use_unsure=use_unsure,
             eval_model=self._eval_model,
+            **(self._question_kwargs or {}),
         )
         query = self._base_query.model_copy()
         query.query = qa_prompt
@@ -305,11 +306,14 @@ def __init__(
         self,
         *args,
         labbench_dataset: str = DEFAULT_LABBENCH_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)
+        train_df, eval_df = read_litqa_v2_from_hub(
+            labbench_dataset, **(read_data_kwargs or {})
+        )
         split = LitQAv2TaskSplit(split)
         if split == LitQAv2TaskSplit.TRAIN:
             self.data = train_df

paperqa/litqa.py

@@ -8,7 +8,7 @@
 from ast import literal_eval
 from collections.abc import Awaitable, Callable, Mapping, Sequence
 from enum import StrEnum
-from typing import TYPE_CHECKING, Self
+from typing import TYPE_CHECKING, Literal, Self
 
 try:
     from ldp.utils import discounted_returns
@@ -92,6 +92,7 @@ def make_mc_options(
 
 DEFAULT_EVAL_MODEL_NAME = "gpt-4-turbo-2024-04-09"
 DEFAULT_REWARD_MAPPING = {"correct": 1.0, "unsure": 0.1, "incorrect": -1.0}
+SEED_USING_QUESTION: Literal["SEED_USING_QUESTION"] = "SEED_USING_QUESTION"  # Sentinel
 
 
 class LitQAEvaluation(StrEnum):
@@ -161,7 +162,7 @@ def from_question(
         question: str,
         use_unsure: bool = True,
         eval_model: LLMModel | str = DEFAULT_EVAL_MODEL_NAME,
-        seed: int | None = None,
+        seed: int | Literal["SEED_USING_QUESTION"] | None = None,
     ) -> tuple[str, Callable[[PQASession | str], Awaitable[LitQAEvaluation]]]:
         """
         Create a LitQA question and an answer-to-evaluation function.
@@ -174,11 +175,15 @@ def from_question(
             eval_model: Evaluation model to use for multiple choice letter extraction
                 from a text answer.
             seed: Optional seed to use in randomization of multiple choice letters.
+                Optionally pass in the string literal "SEED_USING_QUESTION" to hash the
+                input question for the seed.
 
         Returns:
             Two-tuple of created LitQA question, function (that can be thought of as
                 stateless) to use to extract an evaluation result from an answer.
         """
+        if seed == SEED_USING_QUESTION:
+            seed = hash(question)
         text, ideal_answer, unsure_answer, distractor_answers = make_mc_options(
             ideal=ideal,
             distractors=distractors,

tests/test_litqa.py

@@ -3,7 +3,7 @@
 
 import pytest
 
-from paperqa.litqa import LitQAEvaluation, read_litqa_v2_from_hub
+from paperqa.litqa import SEED_USING_QUESTION, LitQAEvaluation, read_litqa_v2_from_hub
 from tests.conftest import VCR_DEFAULT_MATCH_ON
 
 
@@ -140,16 +140,38 @@ def test_consistent_mc_options(self) -> None:
         """Tests that creating multiple evaluations with the same seed results in the same prompt."""
         question, ideal, distractors = self.MEANING_OF_LIFE_QUESTION_IDEAL_DISTRACTORS
 
-        qa_prompt_1, _ = LitQAEvaluation.from_question(
+        qa_prompt_1a, _ = LitQAEvaluation.from_question(
             ideal=ideal, distractors=distractors, question=question, seed=0
         )
-        self._assert_prompt_is_valid(qa_prompt_1, question, ideal, distractors)
+        self._assert_prompt_is_valid(qa_prompt_1a, question, ideal, distractors)
 
-        qa_prompt_2, _ = LitQAEvaluation.from_question(
+        qa_prompt_1b, _ = LitQAEvaluation.from_question(
             ideal=ideal, distractors=distractors, question=question, seed=0
         )
-        self._assert_prompt_is_valid(qa_prompt_1, question, ideal, distractors)
-        assert qa_prompt_1 == qa_prompt_2
+        self._assert_prompt_is_valid(qa_prompt_1b, question, ideal, distractors)
+        assert qa_prompt_1a == qa_prompt_1b, "Same seeding should lead to same prompts"
+
+        qa_prompt_2a, _ = LitQAEvaluation.from_question(
+            ideal=ideal,
+            distractors=distractors,
+            question=question,
+            seed=SEED_USING_QUESTION,
+        )
+        self._assert_prompt_is_valid(qa_prompt_2a, question, ideal, distractors)
+
+        qa_prompt_2b, _ = LitQAEvaluation.from_question(
+            ideal=ideal,
+            distractors=distractors,
+            question=question,
+            seed=SEED_USING_QUESTION,
+        )
+        self._assert_prompt_is_valid(qa_prompt_2b, question, ideal, distractors)
+        assert (
+            qa_prompt_2a == qa_prompt_2b
+        ), "Same seeding strategy should lead to same prompts"
+        assert (
+            qa_prompt_2a != qa_prompt_1a
+        ), "Different seeding strategies should lead to different prompts"
 
     def test_creating_litqa_questions(self) -> None:
         """Test making LitQA eval questions after downloading from Hugging Face Hub."""

tests/test_task.py

@@ -20,7 +20,7 @@
     LitQAv2TaskSplit,
 )
 from paperqa.agents.tools import GenerateAnswer
-from paperqa.litqa import DEFAULT_REWARD_MAPPING, LitQAEvaluation
+from paperqa.litqa import DEFAULT_REWARD_MAPPING, SEED_USING_QUESTION, LitQAEvaluation
 
 
 @pytest.fixture(name="base_query_request")
@@ -103,12 +103,27 @@ async def test___len__(
         expected_length: int,
         base_query_request: QueryRequest,
     ) -> None:
-        task_dataset = LitQAv2TaskDataset(base_query=base_query_request, split=split)
+        task_dataset = LitQAv2TaskDataset(
+            base_query=base_query_request,
+            question_kwargs={"seed": 42},
+            read_data_kwargs={"seed": 42},
+            split=split,
+        )
         assert len(task_dataset) == expected_length
 
         # Now let's check we could use the sources in a validation
         for i in range(len(task_dataset)):
             env = task_dataset.get_new_env_by_idx(i)
+            if i == 0 and split == LitQAv2TaskSplit.TRAIN:
+                # Yes this assertion is somewhat brittle, but it reliably
+                # checks the seeding's behavior so we keep it
+                obs, _ = await env.reset()
+                assert (
+                    "Q: SLC14A1 been identified as a specific marker for endothelial"
+                    " cells in which organ?\n\nOptions:\nA) heart\nB) eye\nC)"
+                    " prostate\nD) Insufficient information to answer this question\nE)"
+                    " liver" in (obs[0].content or "")
+                )
             assert env.sources, "Sources need to be accessible"
             assert isinstance(
                 env.sources, Iterable
@@ -144,6 +159,7 @@ async def test_evaluation(
                         "deleted_dockeys",
                     }
                 ),
+                "question_kwargs": {"seed": SEED_USING_QUESTION},
             },
         )
         # NOTE: set base_query after construction of the TaskConfig. because in