feat: add OLMES variant of HumanEval task (#185)

## PR Checklist
- [X] Use descriptive commit messages.
- [X] Provide tests for your changes.
- [X] Update any related documentation and include any relevant
screenshots.
- [X] Check if changes need to be made to docs (README or any guides in
`/docs/`).

## What type of PR is this? (check all applicable)

- [ ] Refactor
- [X] Feature
- [ ] Bug Fix
- [ ] Optimization
- [X] Documentation Update

## Description

Adds the OLMES variant of the HumanEval

## Added/updated tests?
- [X] Yes
- [ ] No, and this is why: _please replace this line with details on why
tests
      have not been included_
- [ ] I need help with writing tests

Author: tfburns

docs/tasks/HumanEval_OLMES.md

@@ -0,0 +1,20 @@
+# HumanEval_OLMES
+
+````
+NAME = HumanEval_OLMES
+DATASET_PATH = openai/openai_humaneval
+SAMPLE_SPLIT = test
+FEWSHOT_SPLIT = test
+RESPONSE_TYPE = COMPLETION
+METRICS = [CodeCompletionAssertion]
+SUBJECTS = ['no_subject']
+LANGUAGE = <Language.ENG: 'English'>
+````
+
+- Module: `eval_framework.tasks.benchmarks.humaneval`
+
+- File: [src/eval_framework/tasks/benchmarks/humaneval.py](../../src/eval_framework/tasks/benchmarks/humaneval.py) | [View on GitHub](https://github.com/Aleph-Alpha-Research/eval-framework/blob/main/src/eval_framework/tasks/benchmarks/humaneval.py)
+
+- Link to dataset: [https://huggingface.co/datasets/openai/openai_humaneval](https://huggingface.co/datasets/openai/openai_humaneval)
+
+More detailed documentation, with prompt examples and ground truth completions, can be generated with `uv run -m eval_framework.utils.generate_task_docs --add-prompt-examples --only-tasks "HumanEval_OLMES"`.

docs/tasks/README.md

@@ -2,7 +2,7 @@
 
 This directory contains the generated documentation for all benchmark tasks available in the package.
 
-**Total number of tasks: 157**
+**Total number of tasks: 158**
 
 The documentation can be generated or updated with `uv run -m eval_framework.utils.generate_task_docs`.
 
@@ -67,6 +67,7 @@ NOTE: This is an automatically generated file. Any manual modifications will not
 - [HumanEval](HumanEval.md)
 - [HumanEvalBPB](HumanEvalBPB.md)
 - [HumanEvalInstruct](HumanEvalInstruct.md)
+- [HumanEval_OLMES](HumanEval_OLMES.md)
 - [IFEval](IFEval.md)
 - [IFEvalDe](IFEvalDe.md)
 - [IFEvalFiSv](IFEvalFiSv.md)

src/eval_framework/tasks/benchmarks/humaneval.py

@@ -99,6 +99,28 @@ def _get_possible_completions(self, item: dict[str, Any]) -> list[str] | None:
         return [gt] if gt else None
 
 
+class HumanEval_OLMES(HumanEval):
+    """HumanEval OLMES variant replicating codex_humaneval:3shot::olmo3:n32:v2 from oe_eval.
+
+    Recommended EvalConfig settings for full replication:
+        repeats: 32
+        llm_args: {sampling_params: {temperature: 0.6, top_p: 0.6}}
+    """
+
+    NAME = "Human Eval OLMES"
+
+    def __init__(self, num_fewshot: int = 3) -> None:
+        super().__init__(num_fewshot)
+        self.stop_sequences = ["\nclass", "\nif", "\nprint", "\n#", "\n```", "\n```\n\n", "<|eot_id|>"]
+        self.max_tokens = 1024
+
+    def _get_instruction_text(self, item: dict[str, Any]) -> str:
+        return "```python\n" + item["prompt"]
+
+    def _get_fewshot_target_text(self, item: dict[str, Any]) -> str:
+        return item["canonical_solution"] + "```"
+
+
 class HumanEvalInstruct(HumanEval):
     # See https://github.com/EleutherAI/lm-evaluation-harness/blob/main/lm_eval/tasks/humaneval/humaneval_instruct.yaml
     NAME = "Human Eval Instruct"

src/eval_framework/tasks/task_names.py

@@ -61,6 +61,7 @@ def register_all_tasks() -> None:
     register_lazy_task("eval_framework.tasks.benchmarks.opengptx_eu20.HELLASWAG_EU20_FR")
     register_lazy_task("eval_framework.tasks.benchmarks.humaneval.HumanEval")
     register_lazy_task("eval_framework.tasks.benchmarks.humaneval.HumanEvalBPB")
+    register_lazy_task("eval_framework.tasks.benchmarks.humaneval.HumanEval_OLMES")
     register_lazy_task("eval_framework.tasks.benchmarks.humaneval.HumanEvalInstruct")
     register_lazy_task("eval_framework.tasks.benchmarks.ifeval.IFEval")
     register_lazy_task("eval_framework.tasks.benchmarks.ifeval.IFEvalDe")

tests/tests_eval_framework/tasks/task-prompts-hashes.json

@@ -114,6 +114,8 @@
     "HumanEvalBPB.Llama3Formatter": "57f5a23f0cf320ea2c675c178e9f5bad",
     "HumanEvalInstruct.ConcatFormatter": "100e994d25219d93daa3ace8a8beb730",
     "HumanEvalInstruct.Llama3Formatter": "0c1c4e07c9ecd0445257118bc5cecc09",
+    "HumanEval_OLMES.ConcatFormatter": "43d5d2b350304df54708165cf30e4009",
+    "HumanEval_OLMES.Llama3Formatter": "0595b2b93a403ceb3cfef74baa4dd7bf",
     "IFEval.ConcatFormatter": "b517d9d281cd8d5db2ea72b47bd44314",
     "IFEval.Llama3Formatter": "7739c2862af662f2b146f4eae61ac208",
     "IFEvalDe.ConcatFormatter": "798e567efa346a45b42deff904a40b22",

tests/tests_eval_framework/tasks/test_all_formatters.py

@@ -54,6 +54,7 @@
     "HELLASWAG_DE": {"num_fewshot": 1},
     "HELLASWAG_EU20_DE": {"num_fewshot": 1},
     "HELLASWAG_EU20_FR": {"num_fewshot": 1},
+    "HumanEval_OLMES": {"num_fewshot": 3},
     "InfiniteBench_CodeDebug": {"num_fewshot": 0},
     "InfiniteBench_CodeRun": {"num_fewshot": 0},
     "InfiniteBench_EnDia": {"num_fewshot": 0},

tests/tests_eval_framework/tasks/test_humaneval.py

@@ -1,6 +1,6 @@
 import pytest
 
-from eval_framework.tasks.benchmarks.humaneval import HumanEval, HumanEvalInstruct
+from eval_framework.tasks.benchmarks.humaneval import HumanEval, HumanEval_OLMES, HumanEvalInstruct
 from eval_framework.tasks.utils import run_python_code
 from tests.tests_eval_framework.utils import DatasetPatcher
 
@@ -24,6 +24,49 @@ def test_code_is_executed(self, human_eval_task: HumanEval) -> None:
         assert i == 9
 
 
+class TestHumanEvalOLMES:
+    @pytest.fixture
+    def human_eval_olmes_task(self) -> HumanEval_OLMES:
+        with DatasetPatcher(HumanEval_OLMES, num_fewshot=3) as patched_task:
+            return patched_task
+
+    def test_code_is_executed(self, human_eval_olmes_task: HumanEval_OLMES) -> None:
+        assert len(human_eval_olmes_task.SUBJECTS) > 0
+        subject = human_eval_olmes_task.SUBJECTS[0]
+        human_eval_olmes_task._load_dataset(subject)
+        i = 0
+        for i, item in enumerate(human_eval_olmes_task.dataset[human_eval_olmes_task.SAMPLE_SPLIT][:10]):
+            item["subject"] = subject
+            sample = human_eval_olmes_task._create_samples(item, i, subject)[0]
+            formatted_code = human_eval_olmes_task.post_process_generated_completion(item["canonical_solution"], sample)
+            assert run_python_code(formatted_code).endswith("True")
+            formatted_code = human_eval_olmes_task.post_process_generated_completion("", sample)
+            assert not run_python_code(formatted_code).endswith("True")
+        assert i == 9
+
+    def test_olmes_settings(self, human_eval_olmes_task: HumanEval_OLMES) -> None:
+        assert human_eval_olmes_task.num_fewshot == 3
+        assert human_eval_olmes_task.max_tokens == 1024
+        assert "\nclass" in human_eval_olmes_task.stop_sequences
+        assert "\nif" in human_eval_olmes_task.stop_sequences
+        assert "\nprint" in human_eval_olmes_task.stop_sequences
+        assert "\n#" in human_eval_olmes_task.stop_sequences
+        assert "\n```" in human_eval_olmes_task.stop_sequences
+        assert human_eval_olmes_task.SAMPLE_SPLIT == "test"
+        assert human_eval_olmes_task.FEWSHOT_SPLIT == "test"
+
+    def test_olmes_prompt_format(self, human_eval_olmes_task: HumanEval_OLMES) -> None:
+        human_eval_olmes_task._load_dataset(human_eval_olmes_task.SUBJECTS[0])
+        item = human_eval_olmes_task.dataset[human_eval_olmes_task.SAMPLE_SPLIT][0]
+        instruction = human_eval_olmes_task._get_instruction_text(item)
+        assert instruction.startswith("```python\n")
+        assert instruction == "```python\n" + item["prompt"]
+
+        fewshot_target = human_eval_olmes_task._get_fewshot_target_text(item)
+        assert fewshot_target.endswith("```")
+        assert fewshot_target == item["canonical_solution"] + "```"
+
+
 class TestHumanEvalInstructCode:
     @pytest.fixture
     def human_eval_task_inst(self) -> HumanEvalInstruct: