[Evals] Add support for LiveAoPS dataset (#101)

* add liveaops

Signed-off-by: SumanthRH <[email protected]>

* fixes

Signed-off-by: SumanthRH <[email protected]>

* lint

Signed-off-by: SumanthRH <[email protected]>

---------

Signed-off-by: SumanthRH <[email protected]>

Author: SumanthRH

skythought/evals/tasks/__init__.py

@@ -7,6 +7,7 @@
 from .base import ConversationType, TaskConfig, TaskHandler
 from .gpqa_diamond.gpqa_diamond_handler import GPQADiamondTaskHandler
 from .gsm8k.gsm8k_handler import GSM8KTaskHandler
+from .liveaops.liveaops_handler import LiveAOPSTaskHandler
 from .livecodebench.livecodebench_handler import LiveCodeBenchTaskHandler
 from .math.math_handler import MathTaskHandler
 from .minervamath.minervamath_handler import MinervaMathTaskHandler
@@ -33,6 +34,7 @@
     "minervamath": MinervaMathTaskHandler,
     "olympiadbench_math": OlympiadBenchMathTaskHandler,
     "omni_math": OMNIMathTaskHandler,
+    "liveaops": LiveAOPSTaskHandler,
 }
 TASK_NAMES_TO_YAML = get_tasks(os.path.dirname(__file__))
 

skythought/evals/tasks/base.py

@@ -1,5 +1,6 @@
 from abc import ABC, abstractmethod
 from typing import Any, Dict, List, Optional
+from urllib.parse import urlparse
 
 import pandas as pd
 import yaml
@@ -14,7 +15,7 @@ class TaskConfig(BaseModel):
     handler: str
     dataset_path: str
     dataset_subset: Optional[str] = None
-    dataset_split: str
+    dataset_split: Optional[str] = None
     dataset_kwargs: Dict[str, Any] = Field(default_factory=dict)
     question_key: str
     # Optional answer key for datasets with a single correct answer
@@ -82,12 +83,28 @@ def make_conversations(
         return conversations
 
     def load_dataset(self, subset=None, split=None, **kwargs) -> HFDataset:
-        dataset = load_dataset(
-            path=self.task_config.dataset_path,
-            name=subset if subset else self.task_config.dataset_subset,
-            split=split if split else self.task_config.dataset_split,
-            **self.task_config.dataset_kwargs,
-        )
+        # check if the path provided is a valid URL
+        parsed = urlparse(self.task_config.dataset_path)
+        if not parsed.scheme:
+            # HF dataset
+            dataset = load_dataset(
+                path=self.task_config.dataset_path,
+                name=subset if subset else self.task_config.dataset_subset,
+                split=split if split else self.task_config.dataset_split,
+                **self.task_config.dataset_kwargs,
+            )
+        else:
+            # Try to load URL
+            # Only JSON supported for now
+            if split is not None or subset is not None:
+                raise ValueError(
+                    "URL-based dataset does not support loading arguments like `split`, `subset`"
+                )
+            # By default, Huggingface will create a DatasetDict object with "train" split
+            dataset = load_dataset("json", data_files=[self.task_config.dataset_path])[
+                "train"
+            ]
+
         # add an index column efficiently with map
         dataset = dataset.map(add_idx_map, with_indices=True)
         return dataset

skythought/evals/tasks/liveaops/liveaops.yaml

@@ -0,0 +1,8 @@
+handler: liveaops
+dataset_path: https://livemathbench.github.io/data/LiveAoPSBench-2024.jsonl
+dataset_subset: null # which subset on huggingface. Not applicable for a URL dataset
+dataset_split: null # Rule based evaluation
+question_key: question
+answer_key: answer
+templating_parameters: 
+  template: "Return your final response within \\boxed{{}}. {question}"

skythought/evals/tasks/liveaops/liveaops_handler.py

@@ -0,0 +1,26 @@
+from skythought.evals.util.math_parsing_util import (
+    extract_answer,
+    math_equal,
+    strip_answer_string,
+)
+
+from ..math.math_handler import MathTaskHandler
+
+
+class LiveAOPSTaskHandler(MathTaskHandler):
+    def generate_prompt(self, problem):
+        return self.task_config.templating_parameters["template"].format(**problem)
+
+    def check_correctness(self, problem, generation):
+        # no preprocessing needed
+        answer = problem[self.task_config.answer_key]
+        pred = extract_answer(generation)
+        pred = strip_answer_string(pred)
+        return math_equal(pred, answer)
+
+    def load_and_filter_dataset(
+        self, start, end, split=None, subset=None, difficulty=None
+    ):
+        assert difficulty is None, "LiveAOPS does not support `difficulty` argument"
+        dataset = self.load_dataset(subset=subset, split=split).to_pandas()
+        return dataset.iloc[start:end] if end > 0 else dataset.iloc[start:]

skythought/evals/tasks/omni_math/omni_math.yaml

@@ -1,4 +1,4 @@
-handler: math
+handler: omni_math
 dataset_path: "KbsdJames/Omni-MATH"  # repo ID in huggingface
 dataset_subset: null # which subset on huggingface
 dataset_split: test_rule_based # Rule based evaluation