[Examples][Fix] Minor cleanup for the Mini-SWE-Agent example

skyrl-train/docs/examples/mini_swe_agent.rst

@@ -56,10 +56,10 @@ By running this workflow as a Ray task, we are also able to scale up generation
         model = get_model(litellm_model_name, sweagent_config.get("model", {}))
         error = None
         try:
-        env = get_sb_environment(sweagent_config, instance, data_source)
-        agent = DefaultAgent(model, env, **sweagent_config.get("agent", {}))
-        exit_status, model_patch = agent.run(instance["problem_statement"])
-        eval_result = evaluate_trajectory(instance, model_patch, sweagent_config, data_source)
+            env = get_sb_environment(sweagent_config, instance, data_source)
+            agent = DefaultAgent(model, env, **sweagent_config.get("agent", {}))
+            exit_status, model_patch = agent.run(instance["problem_statement"])
+            eval_result = evaluate_trajectory(instance, model_patch, sweagent_config, data_source)
         except Exception as e:
             error = str(e)
         return agent.messages, eval_result, error
@@ -90,7 +90,7 @@ Training
 
 Prerequisites: Ensure that you have the required environment backend installed for generating trajectories with Mini-SWE-Agent. By default, we use `Podman <https://podman.io/docs>`_. This can be modified in :code_link:`examples/mini_swe_agent/swebench.yaml` 
 
-We provide two example scripts: One for Qwen3-8B model and another for the `Qwen/Qwen3-Coder-30B-A3B-Instruct <https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct>` model. While the first script for Qwen3-8B requires a single 8xH100 node, the script for the 30B model requires 2 8xH100 nodes for training.
+We provide two example scripts: One for Qwen3-8B model and another for the `Qwen/Qwen3-Coder-30B-A3B-Instruct <https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct>`_ model. While the first script for Qwen3-8B requires a single 8xH100 node, the script for the 30B model requires 2 8xH100 nodes for training.
 
 .. code-block:: bash
 

skyrl-train/examples/mini_swe_agent/mini_swe_generator.py

@@ -123,10 +123,9 @@ async def minisweagent_agent_loop(
     ) -> Tuple[List[int], float, str, List[int], List[int], Optional[List[int]]]:
 
         sweagent_config = yaml.safe_load(get_config_path(self.generator_cfg.miniswe_config_path).read_text())
-        instance: Dict[str, Dict[str, Any]] = env_extras["instance"]
         # NOTE (sumanthrh): Input `prompt` is not used here because mini-swe-agent uses a similar entry from the `instance` obj
         messages, reward, error = await init_and_run.remote(
-            instance,
+            env_extras["instance"],
             self.litellm_model_name,
             sweagent_config,
             self.generator_cfg,
@@ -136,7 +135,7 @@ async def minisweagent_agent_loop(
         if not len(messages):
             return None, None, None, None, None, None
 
-        # TODO (sumanthrh):This is currently hardcoded for SWEBench with 2 initial messages (system and user).
+        # TODO (sumanthrh): This is currently hardcoded for SWEBench with 2 initial messages (system and user).
         response_messages = messages[2:]
 
         for message in messages[:2]:

skyrl-train/examples/mini_swe_agent/mini_swe_utils.py

@@ -53,9 +53,8 @@ def get_docker_image_name(instance: dict, data_source: str) -> str:
 def evaluate_trajectory(
     instance: Dict[str, Any], model_patch: str, sweagent_config: dict, data_source: str
 ) -> MiniSWEEvaluationResult:
-    instance_id = instance["instance_id"]
 
-    ret = MiniSWEEvaluationResult(instance_id=instance_id, resolved=False, eval_error=None)
+    ret = MiniSWEEvaluationResult(instance_id=instance["instance_id"], resolved=False, eval_error=None)
 
     env = None
     try:

skyrl-train/examples/mini_swe_agent/run_mini_swe_30B.sh

@@ -9,7 +9,9 @@ set -x
 DATA_DIR="$DATA/data/swe_gym_subset"
 
 CKPT_PATH="$DATA/ckpts/llm_mini_swe"
-# save trajectories here for debugging.
+
+# Save trajectories here for debugging.
+# NOTE: For a multi-node cluster, ensure that this is on NFS so that you can save all trajectories in the same path
 MINISWE_TRAJ_DIR="$HOME/mini_swe_agent_trajs_32B"
 
 NUM_GPUS=8

skyrl-train/examples/mini_swe_agent/run_mini_swe_8B.sh

@@ -7,7 +7,9 @@ set -x
 
 DATA_DIR="$HOME/data/swe_gym_subset"
 CKPT_PATH="$HOME/ckpts/llm_mini_swe"
-# save trajectories here
+
+# Save trajectories here for debugging
+# NOTE: For a multi-node cluster, ensure that this is on NFS so that you can save all trajectories in the same path
 MINISWE_TRAJ_DIR="$HOME/mini_swe_agent_trajs"
 
 NUM_GPUS=8