Update SWE-bench instructions, lint (#20)

programmer/evaluate.py

@@ -3,7 +3,7 @@
 import weave
 
 from agent import AgentState
-from config import agent
+from config import agent_4o_basic
 
 # Need to serialize AgentState as json for now since we can't save weave Objects
 # in Dataset set.
@@ -40,7 +40,7 @@ def final_answer_substr(expected_substr: str, model_output: str):
 
 @weave.op()
 def model_agent_bridge(state: str):
-    return agent.run(AgentState(**json.loads(state))).model_dump_json()
+    return agent_4o_basic.run(AgentState(**json.loads(state))).model_dump_json()
 
 
 if __name__ == "__main__":

programmer/swebench/README.md

@@ -1,5 +1,12 @@
 # SWE Bench programmer evaluation
 
+This is a custom setup to run fast SWE-bench evals on programmer. The steps are:
+- serve swebench docker containers from a remote machine
+  - setup an x86 machine (I use a gcp e2-standard-32)
+  - build the swebench instance images. For SWE-bench_Verified this builds about 550 images.
+  - run [containerserver](../containerserver/README.md) on the machine. containerserver serves an HTTP interface into the Docker containers.
+- on your local machine, run python -m programmer.swebench.run_instance or python -m programmer.swebench.evaluate
+
 ## Build SWE-bench images
 
 First do setup (below) then run this command to build all the images. --cache_level instance tells the script not to delete the instance images, which are what we want to use with container-manager.
@@ -13,6 +20,10 @@ python -m swebench.harness.run_evaluation \
      --cache_level instance
 ```
 
+## Run containerserver
+
+See [containerserver](../containerserver/README.md) for setup and running containerserver.
+
 
 ## remote machine setup instructions on gcp VM ubuntu 20.04
 

programmer/swebench/evaluate.py

@@ -7,14 +7,7 @@
 from .swebench_model import SWEBenchProgrammerModel
 from .score import score_swebench
 from ..agent import Agent
-from ..config import SYSTEM_MESSAGE
-from ..tools import (
-    list_files,
-    run_command,
-    view_image,
-    read_lines_from_file,
-    replace_lines_in_file,
-)
+from ..config import agent_4o_basic
 
 
 def load_raw_dataset(name: str, split: str):
@@ -62,22 +55,11 @@ def main():
     # ds = load_weave_dataset("SWE-bench_Verified", "test", instance_ids=instance_ids)
     ds = load_weave_dataset("SWE-bench_Verified", "test", limit=50, shuffle_seed=42)
     eval = weave.Evaluation(
-        name="SWE-bench_Verified", dataset=ds, scorers=[score_swebench], trials=5
+        name="SWE-bench_Verified", dataset=ds, scorers=[score_swebench], trials=1
     )
 
     model = SWEBenchProgrammerModel(
-        agent=Agent(
-            model_name="gpt-4o-2024-08-06",
-            temperature=0.7,
-            system_message=SYSTEM_MESSAGE,
-            tools=[
-                list_files,
-                run_command,
-                view_image,
-                read_lines_from_file,
-                replace_lines_in_file,
-            ],
-        ),
+        agent=agent_4o_basic,
         max_runtime_seconds=180,
     )
     res = asyncio.run(eval.evaluate(model))

programmer/swebench/run_instance.py

@@ -11,7 +11,7 @@
 
 from ..swebench.swebench_model import SWEBenchProgrammerModel
 from ..swebench.score import score_swebench
-from ..config import agent_replace
+from ..config import agent_4o_basic
 
 
 def main():
@@ -42,7 +42,7 @@ def main():
     print("SOLUTION\n", instance["patch"])
     print()
 
-    model = SWEBenchProgrammerModel(agent=agent_replace)
+    model = SWEBenchProgrammerModel(agent=agent_4o_basic)
     model_output = model.predict(instance)
     score = score_swebench(instance, model_output["answer"])
     print("SCORE\n", score)