Add GPQA (#42)

---------

Co-authored-by: Nathan Habib <[email protected]>

src/lighteval/tasks/lighteval_task.py

@@ -279,7 +279,7 @@ def fewshot_docs(self) -> list[Doc]:
             self._fewshot_docs = []
 
             # If we have no available few shot split, the few shot data is the eval data!
-            if self.fewshot_split is None:
+            if self.fewshot_split in [None, [None]]:
                 self._fewshot_docs = self._get_docs_from_split(self.evaluation_split, few_shots=True)
             else:  # Normal case
                 self._fewshot_docs = self._get_docs_from_split(self.fewshot_split, few_shots=True)

src/lighteval/tasks/tasks_prompt_formatting.py

@@ -440,6 +440,26 @@ def ethics_virtue(line, task_name: str = None):
     )
 
 
+def gpqa(line, task_name: str = None):
+    gold_index = random.randint(0, 3)
+    choices = [line["Incorrect Answer 1"], line["Incorrect Answer 2"], line["Incorrect Answer 3"]]
+    choices.insert(gold_index, line["Correct Answer"])
+
+    instruction = "Select the correct answer to the following questions.\n\n"
+
+    query = f"Question: {line['Question']}\n"
+    query += "".join([f"{key}. {choice}\n" for key, choice in zip(LETTER_INDICES, choices)])
+    query += "Answer: "
+
+    return Doc(
+        task_name=task_name,
+        query=f"{instruction}{query}",
+        choices=LETTER_INDICES[: len(choices)],
+        gold_index=gold_index,
+        instruction=instruction,
+    )
+
+
 def gsm8k(line, task_name: str = None):
     # Has special analysis in metric for number decomposiition
     return Doc(

src/lighteval/tasks/tasks_table.jsonl

@@ -359,6 +359,7 @@
 {"name":"glue:stsb","suite":["lighteval","glue"],"prompt_function":"stsb","hf_repo":"glue","hf_subset":"stsb","hf_avail_splits":["test","train","validation"],"evaluation_splits":["validation"],"few_shots_split":null,"few_shots_select":null,"generation_size":1,"metric":["loglikelihood_acc_single_token"],"stop_sequence":["\n"],"output_regex":null,"frozen":false}
 {"name":"glue:wnli","suite":["lighteval","glue"],"prompt_function":"wnli","hf_repo":"glue","hf_subset":"wnli","hf_avail_splits":["test","train","validation"],"evaluation_splits":["validation"],"few_shots_split":null,"few_shots_select":null,"generation_size":1,"metric":["loglikelihood_acc_single_token"],"stop_sequence":["\n"],"output_regex":null,"frozen":false}
 {"name":"goal_step_wikihow","suite":["bigbench","bigbench_json"],"prompt_function":"bigbench","hf_repo":"bigbench","hf_subset":"goal_step_wikihow","hf_avail_splits":["default","train","validation"],"evaluation_splits":["default"],"few_shots_split":null,"few_shots_select":null,"generation_size":1,"metric":["loglikelihood_acc"],"stop_sequence":["\n"],"output_regex":null,"frozen":false}
+{"name":"gpqa","suite":["lighteval"],"prompt_function":"gpqa","hf_repo":"Idavidrein/gpqa","hf_subset":"gpqa_main","hf_avail_splits":["train"],"evaluation_splits":["train"],"few_shots_split":null,"few_shots_select":"random_sampling","generation_size":1,"metric":["loglikelihood_acc_single_token"],"stop_sequence":["\n"],"output_regex":null,"frozen":false}
 {"name":"gre_reading_comprehension","suite":["bigbench","bigbench_json"],"prompt_function":"bigbench","hf_repo":"bigbench","hf_subset":"gre_reading_comprehension","hf_avail_splits":["default","train","validation"],"evaluation_splits":["default"],"few_shots_split":null,"few_shots_select":null,"generation_size":1,"metric":["loglikelihood_acc"],"stop_sequence":["\n"],"output_regex":null,"frozen":false}
 {"name":"gsm8k","suite":["lighteval"],"prompt_function":"gsm8k","hf_repo":"gsm8k","hf_subset":"main","hf_avail_splits":["train","test"],"evaluation_splits":["test"],"few_shots_split":null,"few_shots_select":"random_sampling_from_train","generation_size":256,"metric":["quasi_exact_match_gsm8k"],"stop_sequence":[":","Question:", "Question"],"output_regex":null,"frozen":false}
 {"name":"headqa:en","suite":["lighteval","headqa"],"prompt_function":"headqa","hf_repo":"lighteval/headqa_harness","hf_subset":"en","hf_avail_splits":["train","test","validation"],"evaluation_splits":["test"],"few_shots_split":null,"few_shots_select":null,"generation_size":-1,"metric":["loglikelihood_acc","loglikelihood_acc_norm_nospace"],"stop_sequence":["\n"],"output_regex":null,"frozen":false}