[eval] Support SWE-Bench Multimodal (#7122)

Co-authored-by: openhands <openhands@all-hands.dev>

evaluation/benchmarks/swe_bench/README.md

@@ -2,7 +2,9 @@
 
 This folder contains the evaluation harness that we built on top of the original [SWE-Bench benchmark](https://www.swebench.com/) ([paper](https://arxiv.org/abs/2310.06770)).
 
-**UPDATE (2/18/2025): We now support running SWE-Gym using the same evaluation harness here. For more details, checkout [this README](./SWE-Gym.md).
+**UPDATE (03/27/2025): We now support SWE-Bench multimodal evaluation! Simply use "princeton-nlp/SWE-bench_Multimodal" as the dataset name in the `run_infer.sh` script to evaluate on multimodal instances.**
+
+**UPDATE (2/18/2025): We now support running SWE-Gym using the same evaluation harness here. For more details, checkout [this README](./SWE-Gym.md).**
 
 **UPDATE (7/1/2024): We now support the official SWE-Bench dockerized evaluation as announced [here](https://github.com/princeton-nlp/SWE-bench/blob/main/docs/20240627_docker/README.md).**
 
@@ -62,7 +64,7 @@ in order to use `eval_limit`, you must also set `agent`.
 default, it is set to 60.
 - `num_workers`, e.g. `3`, is the number of parallel workers to run the evaluation. By
 default, it is set to 1.
-- `dataset`, a huggingface dataset name. e.g. `princeton-nlp/SWE-bench`, `princeton-nlp/SWE-bench_Lite`, or `princeton-nlp/SWE-bench_Verified`, specifies which dataset to evaluate on.
+- `dataset`, a huggingface dataset name. e.g. `princeton-nlp/SWE-bench`, `princeton-nlp/SWE-bench_Lite`, `princeton-nlp/SWE-bench_Verified`, or `princeton-nlp/SWE-bench_Multimodal`, specifies which dataset to evaluate on.
 - `dataset_split`, split for the huggingface dataset. e.g., `test`, `dev`. Default to `test`.
 
 > [!CAUTION]
@@ -82,6 +84,13 @@ then your command would be:
 ./evaluation/benchmarks/swe_bench/scripts/run_infer.sh llm.eval_gpt4_1106_preview HEAD CodeActAgent 10
 ```
 
+For multimodal evaluation, you can use:
+
+```bash
+# Example for running multimodal SWE-Bench evaluation
+./evaluation/benchmarks/swe_bench/scripts/run_infer.sh llm.eval_gpt4_vision HEAD CodeActAgent 10 100 1 princeton-nlp/SWE-bench_Multimodal test
+```
+
 ### Running in parallel with RemoteRuntime
 
 OpenHands Remote Runtime is currently in beta (read [here](https://runtime.all-hands.dev/) for more details), it allows you to run rollout in parallel in the cloud, so you don't need a powerful machine to run evaluation.

evaluation/benchmarks/swe_bench/run_infer.py

@@ -58,7 +58,7 @@ def _get_swebench_workspace_dir_name(instance: pd.Series) -> str:
     return f'{instance.repo}__{instance.version}'.replace('/', '__')
 
 
-def get_instruction(instance: pd.Series, metadata: EvalMetadata):
+def get_instruction(instance: pd.Series, metadata: EvalMetadata) -> MessageAction:
     workspace_dir_name = _get_swebench_workspace_dir_name(instance)
     instruction = f"""
 <uploaded_files>
@@ -114,12 +114,20 @@ Be thorough in your exploration, testing, and reasoning. It's fine if your think
 """
 
     if RUN_WITH_BROWSING:
-        instruction += """
+        instruction += (
-<IMPORTANT!>
+            '<IMPORTANT!>\n'
-You SHOULD NEVER attempt to browse the web.
+            'You SHOULD NEVER attempt to browse the web. '
-</IMPORTANT!>
+            '</IMPORTANT!>\n'
-"""
+        )
-    return instruction
+
+    if 'image_assets' in instance:
+        assets = instance['image_assets']
+        assert (
+            'problem_statement' in assets
+        ), 'problem_statement is required in image_assets'
+        image_urls = assets['problem_statement']
+        return MessageAction(content=instruction, image_urls=image_urls)
+    return MessageAction(content=instruction)
 
 
 # TODO: migrate all swe-bench docker to ghcr.io/openhands
@@ -129,14 +137,18 @@ DEFAULT_DOCKER_IMAGE_PREFIX = os.environ.get(
 logger.info(f'Default docker image prefix: {DEFAULT_DOCKER_IMAGE_PREFIX}')
 
 
-def get_instance_docker_image(instance_id: str, official_image: bool = False) -> str:
+def get_instance_docker_image(
-    if official_image:
+    instance_id: str,
+    swebench_official_image: bool = False,
+) -> str:
+    if swebench_official_image:
         # Official SWE-Bench image
         # swebench/sweb.eval.x86_64.django_1776_django-11333:v1
         docker_image_prefix = 'docker.io/swebench/'
         repo, name = instance_id.split('__')
-        image_name = f'sweb.eval.x86_64.{repo}_1776_{name}:latest'
+        image_name = f'swebench/sweb.eval.x86_64.{repo}_1776_{name}:latest'
-        logger.warning(f'Using official SWE-Bench image: {image_name}')
+        logger.info(f'Using official SWE-Bench image: {image_name}')
+        return image_name
     else:
         # OpenHands version of the image
         docker_image_prefix = DEFAULT_DOCKER_IMAGE_PREFIX
@@ -144,7 +156,7 @@ def get_instance_docker_image(instance_id: str, official_image: bool = False) ->
         image_name = image_name.replace(
             '__', '_s_'
         )  # to comply with docker image naming convention
-    return (docker_image_prefix.rstrip('/') + '/' + image_name).lower()
+        return (docker_image_prefix.rstrip('/') + '/' + image_name).lower()
 
 
 def get_config(
@@ -152,12 +164,13 @@ def get_config(
     metadata: EvalMetadata,
 ) -> AppConfig:
     # We use a different instance image for the each instance of swe-bench eval
-    use_official_image = bool(
+    use_swebench_official_image = bool(
         ('verified' in metadata.dataset.lower() or 'lite' in metadata.dataset.lower())
         and 'swe-gym' not in metadata.dataset.lower()
     )
     base_container_image = get_instance_docker_image(
-        instance['instance_id'], use_official_image
+        instance['instance_id'],
+        swebench_official_image=use_swebench_official_image,
     )
     logger.info(
         f'Using instance container image: {base_container_image}. '
@@ -493,13 +506,13 @@ def process_instance(
     try:
         initialize_runtime(runtime, instance)
 
-        instruction = get_instruction(instance, metadata)
+        message_action = get_instruction(instance, metadata)
 
         # Here's how you can run the agent (similar to the `main` function) and get the final task state
         state: State | None = asyncio.run(
             run_controller(
                 config=config,
-                initial_user_action=MessageAction(content=instruction),
+                initial_user_action=message_action,
                 runtime=runtime,
                 fake_user_response_fn=AGENT_CLS_TO_FAKE_USER_RESPONSE_FN[
                     metadata.agent_class
@@ -539,6 +552,11 @@ def process_instance(
     metrics = get_metrics(state)
 
     # Save the output
+    instruction = message_action.content
+    if message_action.image_urls:
+        instruction += (
+            '\n\n<image_urls>' + '\n'.join(message_action.image_urls) + '</image_urls>'
+        )
     output = EvalOutput(
         instance_id=instance.instance_id,
         instruction=instruction,