Use bespokelabs-sandbox to run code in sandboxes

.github/workflows/lint-and-test.yaml

@@ -43,7 +43,7 @@ jobs:
       - name: Install dependencies
         run: |
           pip install poetry
-          poetry install --with dev --extras "vllm code_execution"
+          poetry install --with dev --extras "vllm"
 
       - name: Run ruff
         run: |

examples/code-execution/chart-generation/charts.py

@@ -108,16 +108,11 @@ def code_output(self, row, execution_output):
                     tar.extractall(path=temp_dir)
 
                 # read the chart image
-                # it is in workspace/chart.png for 'docker' backend and chart.png for 'multiprocessing' backend
                 try:
-                    with open(os.path.join(temp_dir, "workspace/chart.png"), "rb") as f:
+                    with open(os.path.join(temp_dir, "chart.png"), "rb") as f:
                         row["chart_image"] = f.read()
                 except Exception:
-                    try:
-                        with open(os.path.join(temp_dir, "chart.png"), "rb") as f:
-                            row["chart_image"] = f.read()
-                    except Exception:
-                        row["chart_image"] = None
+                    row["chart_image"] = None
 
         return row
 
@@ -126,7 +121,7 @@ def code_output(self, row, execution_output):
     # This section demonstrates how to use the ChartCodeExecutor
 
     # Initialize executor with the multiprocessing backend (runs code in a separate process)
-    executor = ChartCodeExecutor(backend="multiprocessing")  # or ray, docker (with matplotlib image)
+    executor = ChartCodeExecutor(backend="local")  # or docker, e2b, modal, daytona
 
     # Create a simple dataset with one example
     dataset = Dataset.from_list([{"id": 1}])

examples/code-execution/code-verification/code_verify.py

@@ -1,4 +1,4 @@
-"""This example demonstrates how to use Curator's code execution functionality with Ray backend for executing code on large datasets.
+"""This example demonstrates how to use Curator's code execution functionality for executing code on large datasets.
 
 It shows how to verify code solutions against expected outputs.
 
@@ -60,7 +60,7 @@ def code_output(self, row, execution_output):
 
 
 if __name__ == "__main__":
-    executor = TACOCodeExecutor(backend="ray")  # or multiprocessing, docker, e2b
+    executor = TACOCodeExecutor(backend="local")  # or docker, e2b, modal, daytona
     dataset = load_dataset("bespokelabs/sky-t1-taco-test-rejection-sampled-shreyas")
     execution_output = executor(dataset["train"].select(range(10)))
 

examples/code-execution/math-animation/execute_code.py

@@ -75,8 +75,8 @@ def code_output(self, input: Dict, output) -> Dict:
                     with tarfile.open(fileobj=bytes_io) as tar:
                         tar.extractall(path=temp_dir)
 
-                    # get the video file
-                    video_file = os.path.join(temp_dir, "workspace/media/videos/temp_animation/720p30/video.mp4")
+                    # get the video file (path relative to sandbox workspace)
+                    video_file = os.path.join(temp_dir, "media/videos/temp_animation/720p30/video.mp4")
                     if os.path.exists(video_file):
                         # Copy the video file to the output dataset
                         input["video"] = open(video_file, "rb").read()
@@ -111,7 +111,7 @@ def execute_manim_code(dataset_name="pimpalgaonkar/manim_codes_10k", output_data
     executor = ManimCodeExecutor(
         backend="docker",
         backend_params={
-            "docker_image": "manimcommunity/manim:latest",
+            "image": "manimcommunity/manim:latest",
         },
     )
 

pyproject.toml

@@ -35,17 +35,14 @@ aiofiles = ">=22.0,<26.0"
 anthropic = "^0.84.0"
 vertexai = "1.71.1"
 vllm = { version = "^0.6.3", optional = true }
-aiodocker = { version = "^0.24.0", optional = true }
-ray = { version = "^2.41.0", optional = true }
-e2b-code-interpreter = { version = "1.0.3", optional = true }
+bespokelabs-sandbox = "0.1.1"
 tinker = { version = "^0.8.0", optional = true, python = ">=3.11" }
 
 posthog = "^3.11.0"
 mistralai = "^1.5.1"
 
 [tool.poetry.extras]
 vllm = ["vllm"]
-code_execution = ["ray", "aiodocker", "e2b-code-interpreter"]
 finetune = ["tinker"]
 
 [tool.poetry.group.dev.dependencies]

src/bespokelabs/curator/code_executor/code_execution_backend/_factory.py

@@ -1,31 +1,35 @@
+"""Factory for creating code execution backends."""
+
 from typing import Optional
 
 from bespokelabs.curator.code_executor.types import CodeExecutionBackendConfig
 
+# Map old backend names to sandbox backend names
+_BACKEND_ALIASES = {
+    "multiprocessing": "local",
+}
+
 
 class _CodeExecutionBackendFactory:
-    @classmethod
-    def create(cls, backend: str, backend_params: Optional[CodeExecutionBackendConfig] = None):
-        if backend_params is None:
-            backend_params = {}
+    """Factory for creating code execution backends."""
 
-        if backend == "multiprocessing":
-            from bespokelabs.curator.code_executor.code_execution_backend.multiprocessing_backend import MultiprocessingCodeExecutionBackend
+    @classmethod
+    def create(cls, backend: str, backend_params: Optional[dict] = None):
+        """Create a code execution backend.
 
-            _code_executor = MultiprocessingCodeExecutionBackend(CodeExecutionBackendConfig(**backend_params))
-        elif backend == "docker":
-            from bespokelabs.curator.code_executor.code_execution_backend.docker_backend import DockerCodeExecutionBackend
+        Args:
+            backend: Backend name (e.g. "local", "docker", "e2b", "modal", "daytona")
+            backend_params: Configuration parameters for the backend
 
-            _code_executor = DockerCodeExecutionBackend(CodeExecutionBackendConfig(**backend_params))
-        elif backend == "ray":
-            from bespokelabs.curator.code_executor.code_execution_backend.ray_backend import RayCodeExecutionBackend
+        Returns:
+            SandboxCodeExecutionBackend instance
+        """
+        if backend_params is None:
+            backend_params = {}
 
-            _code_executor = RayCodeExecutionBackend(CodeExecutionBackendConfig(**backend_params))
-        elif backend == "e2b":
-            from bespokelabs.curator.code_executor.code_execution_backend.e2b_backend import E2BCodeExecutionBackend
+        backend_name = _BACKEND_ALIASES.get(backend, backend)
+        config = CodeExecutionBackendConfig(**backend_params)
 
-            _code_executor = E2BCodeExecutionBackend(CodeExecutionBackendConfig(**backend_params))
-        else:
-            raise ValueError(f"Unsupported backend: {backend}")
+        from bespokelabs.curator.code_executor.code_execution_backend.sandbox_backend import SandboxCodeExecutionBackend
 
-        return _code_executor
+        return SandboxCodeExecutionBackend(config, backend_name=backend_name)

src/bespokelabs/curator/code_executor/code_execution_backend/base_backend.py

@@ -2,11 +2,9 @@
 
 import asyncio
 import glob
-import io
 import json
 import os
 import resource
-import tarfile
 import time
 from abc import abstractmethod
 from typing import TYPE_CHECKING, List, Optional
@@ -656,47 +654,3 @@ async def append_execution_response(self, data: dict, filename: str) -> None:
         async with aiofiles.open(filename, "a") as f:
             await f.write(json_string + "\n")
         logger.debug(f"Successfully appended response to {filename}")
-
-    @classmethod
-    def _create_temp_file(cls, content: str, execution_directory: str) -> str:
-        """Create a temporary file with the given content.
-
-        Args:
-            content: Content to write to temp file
-            execution_directory: Directory to create the temp file in
-
-        Returns:
-            Path to the created temp file
-        """
-        # create execution directory if it doesn't exist
-        os.makedirs(execution_directory, exist_ok=True)
-
-        temp_file_path = os.path.join(execution_directory, "program.py")
-        with open(temp_file_path, "w", encoding="utf-8") as temp_file:
-            temp_file.write(content)
-        return temp_file_path
-
-    @classmethod
-    def _get_created_files(cls, program_dir: str) -> bytes:
-        """Get any files created during code execution.
-
-        Args:
-            program_dir: Directory containing the executed program and created files
-
-        Returns:
-            Bytes containing a zip archive of any created files
-        """
-        # Create a BytesIO object to store the zip data
-        tar_buffer = io.BytesIO()
-
-        # Create a zip archive containing all files in program_dir
-        with tarfile.open(fileobj=tar_buffer, mode="w") as tar:
-            # Add all files in program_dir to the archive
-            for filename in os.listdir(program_dir):
-                file_path = os.path.join(program_dir, filename)
-                if os.path.isfile(file_path):
-                    tar.add(file_path, arcname=filename)
-
-        # Get the bytes from the buffer
-        tar_buffer.seek(0)
-        return str(tar_buffer.getvalue())