Merge pull request rllm-org#156 from MattFisher/cleanup/class_eval

Follow-up: ClassEval

Author: alexandraabbas

README.md

@@ -83,6 +83,14 @@ Inspect supports many model providers including OpenAI, Anthropic, Google, Mistr
    inspect eval inspect_evals/ds1000
    ```
 
+- ### [ClassEval: A Manually-Crafted Benchmark for Evaluating LLMs on Class-level Code Generation](src/inspect_evals/class_eval)
+  Evaluates LLMs on class-level code generation with 100 tasks constructed over 500 person-hours. The study shows that LLMs perform worse on class-level tasks compared to method-level tasks. GPT-4 and GPT-3.5 outperform other models, with holistic generation being the best strategy for them, while incremental generation works better for other models. The study also shows that the performance of LLMs is highly correlated with the number of tokens in the prompt.
+ <sub><sup>Contributed by: [@zhenningdavidliu](https://github.com/zhenningdavidliu)</sub></sup>
+
+   ```bash
+   inspect eval inspect_evals/class_eval
+   ```
+
 ## Assistants
 
 - ### [GAIA: A Benchmark for General AI Assistants](src/inspect_evals/gaia)

src/inspect_evals/_registry.py

@@ -16,6 +16,7 @@
 from .assistant_bench import assistant_bench
 from .bbh import bbh
 from .boolq import boolq
+from .class_eval import class_eval
 from .commonsense_qa import commonsense_qa
 from .cybench import cybench
 from .cybermetric import (

src/inspect_evals/class_eval/README.md

@@ -0,0 +1,66 @@
+# ClassEval
+
+[ClassEval](https://github.com/FudanSELab/ClassEval) is a benchmark for evaluating large language models (LLMs) on class-level code generation tasks. It contains 100 class-level Python code generation tasks, constructed over approximately 500 person-hours.
+
+<!-- Contributors: Automatically Generated -->
+Contributed by [@zhenningdavidliu](https://github.com/zhenningdavidliu)
+<!-- /Contributors: Automatically Generated -->
+
+<!-- Usage: Automatically Generated -->
+## Usage
+
+First, install the `inspect_ai` and `inspect_evals` Python packages with:
+
+```bash
+pip install inspect_ai
+pip install git+https://github.com/UKGovernmentBEIS/inspect_evals
+```
+
+Then, evaluate against one or more models with:
+
+```bash
+inspect eval inspect_evals/class_eval --model openai/gpt-4o
+```
+
+After running evaluations, you can view their logs using the `inspect view` command:
+
+```bash
+inspect view
+```
+
+If you don't want to specify the `--model` each time you run an evaluation, create a `.env` configuration file in your working directory that defines the `INSPECT_EVAL_MODEL` environment variable along with your API key. For example:
+
+```bash
+INSPECT_EVAL_MODEL=anthropic/claude-3-5-sonnet-20240620
+ANTHROPIC_API_KEY=<anthropic-api-key>
+```
+<!-- /Usage: Automatically Generated -->
+
+>[!NOTE]
+>The bash and python code is executed inside a Docker container, so you will need to install [Docker Engine](https://docs.docker.com/engine/install/) in order to run the evaluation.
+>
+> Docker containers are also used in some challenges for auxiliary services. When you run this, the containers are pulled from a public registry. The source code is still included, however, both for your reference and because some files are copied from them into the agent's container at runtime.
+
+<!-- Options: Automatically Generated -->
+## Options
+
+You can control a variety of options from the command line. For example:
+
+```bash
+inspect eval inspect_evals/class_eval --limit 10
+inspect eval inspect_evals/class_eval --max-connections 10
+inspect eval inspect_evals/class_eval --temperature 0.5
+```
+
+See `inspect eval --help` for all available options.
+<!-- /Options: Automatically Generated -->
+
+## Dataset
+
+See https://huggingface.co/datasets/FudanSELab/ClassEval for more information.
+
+
+## Scoring
+For each sample, the LLM-generated code is appended with test cases from the dataset, and the code is executed to check if it passes all the test cases. If all test cases pass, the sample is considered correct, and the score for the sample is 1. Otherwise, the score for the sample is 0.
+
+The final score for the model is the average score across all samples.

src/inspect_evals/class_eval/__init__.py

@@ -1 +1,3 @@
-from .class_eval import class_eval
+from .class_eval import class_eval
+
+__all__ = ["class_eval"]

src/inspect_evals/class_eval/class_eval.py

@@ -11,23 +11,22 @@
 import re
 from typing import Any
 
-from inspect_ai import Task, task, Epochs
+from inspect_ai import Epochs, Task, task
 from inspect_ai.dataset import Sample, hf_dataset
 from inspect_ai.scorer import (
     CORRECT,
     INCORRECT,
     Score,
     Scorer,
     Target,
-    scorer,
     mean,
+    scorer,
     std,
 )
 from inspect_ai.solver import (
-    generate, 
-    Solver, 
-    system_message,
     TaskState,
+    generate,
+    system_message,
 )
 from inspect_ai.util import ExecResult, sandbox
 
@@ -36,22 +35,19 @@
 # Timeout for scoring.
 VERIFY_TIMEOUT = 30
 
+
 @task
 def class_eval(few_shot: int = 1, few_shot_seed: int = 42) -> Task:
     """Inspect Task implementation of ClassEval.
 
     Args:
-        k_shot (int): The number of few shots to include.
-        k_shot_seed (int): The seed for generating few shots.
-        solver (Solver): The solver to use for this evaluation. Defaults to the default solver.
-        instruction_prompt (String): The prompt to prepend to the code problem.
-        scorer (Scorer): The scorer to use for this evaluation. Defaults to the default scorer.
+        few_shot (int): The number of few shots to include.
+        few_shot_seed (int): The seed for generating few shots.
     """
-
     dataset = hf_dataset(
         path="FudanSELab/ClassEval",
         split="test",
-        sample_fields=record_to_sample
+        sample_fields=record_to_sample,
     )
 
     INSTRUCTION = """
@@ -64,23 +60,23 @@ def class_eval(few_shot: int = 1, few_shot_seed: int = 42) -> Task:
         dataset=dataset,
         solver=[system_message(INSTRUCTION), generate()],
         scorer=class_eval_scorer(),
-        epochs = Epochs(few_shot, [f"pass_at_{few_shot}"]),
+        epochs=Epochs(few_shot, [f"pass_at_{few_shot}"]),
         sandbox="docker",
     )
 
 
 @scorer(metrics=[mean(), std()])
 def class_eval_scorer() -> Scorer:
-    '''
-    This is the scorer for class eval. It will first identify the python code in the output of 
-    the model. Then we append the test cases to the code and execute it. If the code passes all the test cases,
-    we return a CORRECT score. Otherwise, we return an INCORRECT score.
-    '''
+    """
+    Scorer for class eval
 
-    async def score(state:TaskState, target: Target) -> Score:
-        
-        result = {}
+    It will first identify the python code in the output of the model.
+    Then we append the test cases to the code and execute it.
+    If the code passes all the test cases, we return a CORRECT score.
+    Otherwise, we return an INCORRECT score.
+    """
 
+    async def score(state: TaskState, target: Target) -> Score:
         generated_code = find_code(state.output.completion)
         code = generated_code + "\n" + state.metadata["test"]
 
@@ -90,7 +86,7 @@ async def score(state:TaskState, target: Target) -> Score:
         explanation += "\n```\n"
 
         try:
-            result = await sandbox().exec(
+            result: ExecResult[str] = await sandbox().exec(
                 cmd=["python", "-c", code],
                 timeout=VERIFY_TIMEOUT,
             )
@@ -100,7 +96,9 @@ async def score(state:TaskState, target: Target) -> Score:
             else:
                 explanation += "Code did not pass all test cases.\n"
                 if result.stderr:
-                    explanation += f"See details below.\n ```python\n {result.stderr} \n ```\n"
+                    explanation += (
+                        f"See details below.\n ```python\n {result.stderr} \n ```\n"
+                    )
         except TimeoutError:
             result = ExecResult(False, 1, "", "Verification timed out.")
             explanation += "Verification timed out."
@@ -126,13 +124,11 @@ def find_code(completion: str) -> str:
 def record_to_sample(
     record: dict[str, Any],
 ) -> Sample:
-    '''
-    Maps class_eval record into inspect sample
-    '''
+    """Convert a record from the dataset to a Sample"""
     return Sample(
-        input = construct_prompt(record),
-        target =  record["solution_code"],
-        id = record["task_id"],
+        input=construct_prompt(record),
+        target=record["solution_code"],
+        id=record["task_id"],
         metadata={
             "task_id": record["task_id"],
             "skeleton": record["skeleton"],
@@ -145,5 +141,5 @@ def record_to_sample(
             "test_classes": record["test_classes"],
             "class_constructor": record["class_constructor"],
             "fields": record["fields"],
-        }
+        },
     )