SWE training sandbox migration

benchmarks/nemotron_3.5_super/eval_container_config.yaml

@@ -11,6 +11,11 @@ config_paths:
 - benchmarks/deepswe/opencode.yaml
 - benchmarks/terminal_bench_2_1/terminus_2.yaml
 - benchmarks/swebench/pro/opencode.yaml
+- benchmarks/reward_profiling/swe_rebench/opencode.yaml
+- benchmarks/reward_profiling/swemer_v1/opencode.yaml
+- benchmarks/reward_profiling/swemer_v2/opencode.yaml
+- benchmarks/reward_profiling/swe_next/opencode.yaml
+- benchmarks/reward_profiling/scale_swe/opencode.yaml
 - responses_api_models/vllm_model/configs/vllm_model.yaml
 - benchmarks/nemotron_3.5_super/sandbox_utils.yaml
 
@@ -116,6 +121,26 @@ swebench_pro_opencode_sandboxed_agent:
   responses_api_agents:
     opencode_sandboxed_agent:
       datasets: []
+swe_rebench_reward_profiling_opencode_sandboxed_agent:
+  responses_api_agents:
+    opencode_sandboxed_agent:
+      datasets: []
+swemer_v1_reward_profiling_opencode_sandboxed_agent:
+  responses_api_agents:
+    opencode_sandboxed_agent:
+      datasets: []
+swemer_v2_reward_profiling_opencode_sandboxed_agent:
+  responses_api_agents:
+    opencode_sandboxed_agent:
+      datasets: []
+swe_next_reward_profiling_opencode_sandboxed_agent:
+  responses_api_agents:
+    opencode_sandboxed_agent:
+      datasets: []
+scale_swe_reward_profiling_opencode_sandboxed_agent:
+  responses_api_agents:
+    opencode_sandboxed_agent:
+      datasets: []
 
 ########################################
 # END Nvidia-specific configs.

benchmarks/nemotron_3.5_super/eval_container_config_with_staged.yaml

@@ -50,4 +50,54 @@ swebench_pro_opencode_sandboxed_agent:
           type: benchmark
           jsonl_fpath: benchmarks/swebench/data/swebench_pro_benchmark.jsonl
           prepare_script: benchmarks/swebench/pro/prepare.py
-          num_repeats: 3
+          num_repeats: 3
+
+swe_rebench_reward_profiling_opencode_sandboxed_agent:
+  responses_api_agents:
+    opencode_sandboxed_agent:
+      datasets:
+        - name: swe_rebench_reward_profiling
+          type: benchmark
+          jsonl_fpath: resources_servers/swe_rebench/data/swe_rebench_training_enriched_with_prompt_template.jsonl
+          prepare_script: benchmarks/reward_profiling/swe_rebench/prepare.py
+          num_repeats: 8
+
+swemer_v1_reward_profiling_opencode_sandboxed_agent:
+  responses_api_agents:
+    opencode_sandboxed_agent:
+      datasets:
+        - name: swemer_v1_reward_profiling
+          type: benchmark
+          jsonl_fpath: resources_servers/swemer_v1/data/swemer_v1_training_enriched_with_prompt_template.jsonl
+          prepare_script: benchmarks/reward_profiling/swemer_v1/prepare.py
+          num_repeats: 8
+
+swemer_v2_reward_profiling_opencode_sandboxed_agent:
+  responses_api_agents:
+    opencode_sandboxed_agent:
+      datasets:
+        - name: swemer_v2_reward_profiling
+          type: benchmark
+          jsonl_fpath: resources_servers/swemer_v2/data/swemer_v2_training_enriched_with_prompt_template.jsonl
+          prepare_script: benchmarks/reward_profiling/swemer_v2/prepare.py
+          num_repeats: 8
+
+swe_next_reward_profiling_opencode_sandboxed_agent:
+  responses_api_agents:
+    opencode_sandboxed_agent:
+      datasets:
+        - name: swe_next_reward_profiling
+          type: benchmark
+          jsonl_fpath: resources_servers/swe_next/data/swe_next_training_with_prompt_template.jsonl
+          prepare_script: benchmarks/reward_profiling/swe_next/prepare.py
+          num_repeats: 8
+
+scale_swe_reward_profiling_opencode_sandboxed_agent:
+  responses_api_agents:
+    opencode_sandboxed_agent:
+      datasets:
+        - name: scale_swe_reward_profiling
+          type: benchmark
+          jsonl_fpath: resources_servers/scale_swe/data/scale_swe_training_with_prompt_template.jsonl
+          prepare_script: benchmarks/reward_profiling/scale_swe/prepare.py
+          num_repeats: 8

benchmarks/reward_profiling/scale_swe/opencode.yaml

@@ -0,0 +1,36 @@
+# Reward profiling: run scale_swe's training tasks through the opencode sandboxed agent (not
+# the golden patch) to collect real rollouts + rewards for a policy checkpoint, 8 repeats per
+# task. See resources_servers/scale_swe/configs/scale_swe.yaml for the base resources server.
+config_paths:
+- responses_api_agents/opencode_sandboxed_agent/configs/opencode_sandboxed_agent.yaml
+- resources_servers/scale_swe/configs/scale_swe.yaml
+
+scale_swe_reward_profiling_resources_server:
+  _inherit_from: scale_swe_resources_server
+
+  resources_servers:
+    scale_swe:
+      allowed_agents: [opencode_sandboxed_agent]
+      evaluation_timeout: 1200
+
+scale_swe_reward_profiling_opencode_sandboxed_agent:
+  _inherit_from: opencode_sandboxed_agent
+  responses_api_agents:
+    opencode_sandboxed_agent:
+      resources_server:
+        type: resources_servers
+        name: scale_swe_reward_profiling_resources_server
+      # 1hr harness execution budget per rollout, not the eval-server's own evaluation_timeout.
+      sandbox_timeout: 3600
+      # Keep every full turn in the rollout for reward profiling -- no OpenCode-side
+      # auto-summarization/pruning of earlier context.
+      opencode_config:
+        compaction:
+          auto: false
+          prune: false
+      datasets:
+        - name: scale_swe_reward_profiling
+          type: benchmark
+          jsonl_fpath: resources_servers/scale_swe/data/scale_swe_training_with_prompt_template.jsonl
+          prepare_script: benchmarks/reward_profiling/scale_swe/prepare.py
+          num_repeats: 8

benchmarks/reward_profiling/scale_swe/prepare.py

@@ -0,0 +1,40 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""No-op prepare script for reward profiling: the jsonl this benchmark points at is already
+fully prepared (responses_create_params with the prompt template applied). `gym eval prepare`
+still requires an importable `prepare_script` module with a `prepare()` function, so this exists
+to satisfy that -- it is never actually called when invoked with
+`+use_cached_prepared_benchmarks=true` (the standard invocation) and the jsonl already exists.
+"""
+
+from pathlib import Path
+
+
+OUTPUT_FPATH = (
+    Path(__file__).parent.parent.parent.parent
+    / "resources_servers"
+    / "scale_swe"
+    / "data"
+    / "scale_swe_training_with_prompt_template.jsonl"
+)
+
+
+def prepare() -> Path:
+    if not OUTPUT_FPATH.exists():
+        raise FileNotFoundError(
+            f"{OUTPUT_FPATH} does not exist. This benchmark expects a pre-built jsonl; there is "
+            "no generation logic here to build one."
+        )
+    return OUTPUT_FPATH

benchmarks/reward_profiling/swe_next/opencode.yaml

@@ -0,0 +1,36 @@
+# Reward profiling: run swe_next's training tasks through the opencode sandboxed agent (not
+# the golden patch) to collect real rollouts + rewards for a policy checkpoint, 8 repeats per
+# task. See resources_servers/swe_next/configs/swe_next.yaml for the base resources server.
+config_paths:
+- responses_api_agents/opencode_sandboxed_agent/configs/opencode_sandboxed_agent.yaml
+- resources_servers/swe_next/configs/swe_next.yaml
+
+swe_next_reward_profiling_resources_server:
+  _inherit_from: swe_next_resources_server
+
+  resources_servers:
+    swe_next:
+      allowed_agents: [opencode_sandboxed_agent]
+      evaluation_timeout: 1200
+
+swe_next_reward_profiling_opencode_sandboxed_agent:
+  _inherit_from: opencode_sandboxed_agent
+  responses_api_agents:
+    opencode_sandboxed_agent:
+      resources_server:
+        type: resources_servers
+        name: swe_next_reward_profiling_resources_server
+      # 1hr harness execution budget per rollout, not the eval-server's own evaluation_timeout.
+      sandbox_timeout: 3600
+      # Keep every full turn in the rollout for reward profiling -- no OpenCode-side
+      # auto-summarization/pruning of earlier context.
+      opencode_config:
+        compaction:
+          auto: false
+          prune: false
+      datasets:
+        - name: swe_next_reward_profiling
+          type: benchmark
+          jsonl_fpath: resources_servers/swe_next/data/swe_next_training_with_prompt_template.jsonl
+          prepare_script: benchmarks/reward_profiling/swe_next/prepare.py
+          num_repeats: 8

benchmarks/reward_profiling/swe_next/prepare.py

@@ -0,0 +1,40 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""No-op prepare script for reward profiling: the jsonl this benchmark points at is already
+fully prepared (responses_create_params with the prompt template applied). `gym eval prepare`
+still requires an importable `prepare_script` module with a `prepare()` function, so this exists
+to satisfy that -- it is never actually called when invoked with
+`+use_cached_prepared_benchmarks=true` (the standard invocation) and the jsonl already exists.
+"""
+
+from pathlib import Path
+
+
+OUTPUT_FPATH = (
+    Path(__file__).parent.parent.parent.parent
+    / "resources_servers"
+    / "swe_next"
+    / "data"
+    / "swe_next_training_with_prompt_template.jsonl"
+)
+
+
+def prepare() -> Path:
+    if not OUTPUT_FPATH.exists():
+        raise FileNotFoundError(
+            f"{OUTPUT_FPATH} does not exist. This benchmark expects a pre-built jsonl; there is "
+            "no generation logic here to build one."
+        )
+    return OUTPUT_FPATH

benchmarks/reward_profiling/swe_rebench/opencode.yaml

@@ -0,0 +1,36 @@
+# Reward profiling: run swe_rebench's training tasks through the opencode sandboxed agent (not
+# the golden patch) to collect real rollouts + rewards for a policy checkpoint, 8 repeats per
+# task. See resources_servers/swe_rebench/configs/swe_rebench.yaml for the base resources server.
+config_paths:
+- responses_api_agents/opencode_sandboxed_agent/configs/opencode_sandboxed_agent.yaml
+- resources_servers/swe_rebench/configs/swe_rebench.yaml
+
+swe_rebench_reward_profiling_resources_server:
+  _inherit_from: swe_rebench_resources_server
+
+  resources_servers:
+    swe_rebench:
+      allowed_agents: [opencode_sandboxed_agent]
+      evaluation_timeout: 1200
+
+swe_rebench_reward_profiling_opencode_sandboxed_agent:
+  _inherit_from: opencode_sandboxed_agent
+  responses_api_agents:
+    opencode_sandboxed_agent:
+      resources_server:
+        type: resources_servers
+        name: swe_rebench_reward_profiling_resources_server
+      # 1hr harness execution budget per rollout, not the eval-server's own evaluation_timeout.
+      sandbox_timeout: 3600
+      # Keep every full turn in the rollout for reward profiling -- no OpenCode-side
+      # auto-summarization/pruning of earlier context.
+      opencode_config:
+        compaction:
+          auto: false
+          prune: false
+      datasets:
+        - name: swe_rebench_reward_profiling
+          type: benchmark
+          jsonl_fpath: resources_servers/swe_rebench/data/swe_rebench_training_enriched_with_prompt_template.jsonl
+          prepare_script: benchmarks/reward_profiling/swe_rebench/prepare.py
+          num_repeats: 8

benchmarks/reward_profiling/swe_rebench/prepare.py

@@ -0,0 +1,40 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""No-op prepare script for reward profiling: the jsonl this benchmark points at is already
+fully prepared (responses_create_params with the prompt template applied). `gym eval prepare`
+still requires an importable `prepare_script` module with a `prepare()` function, so this exists
+to satisfy that -- it is never actually called when invoked with
+`+use_cached_prepared_benchmarks=true` (the standard invocation) and the jsonl already exists.
+"""
+
+from pathlib import Path
+
+
+OUTPUT_FPATH = (
+    Path(__file__).parent.parent.parent.parent
+    / "resources_servers"
+    / "swe_rebench"
+    / "data"
+    / "swe_rebench_training_enriched_with_prompt_template.jsonl"
+)
+
+
+def prepare() -> Path:
+    if not OUTPUT_FPATH.exists():
+        raise FileNotFoundError(
+            f"{OUTPUT_FPATH} does not exist. This benchmark expects a pre-built jsonl; there is "
+            "no generation logic here to build one."
+        )
+    return OUTPUT_FPATH

benchmarks/reward_profiling/swemer_v1/opencode.yaml

@@ -0,0 +1,36 @@
+# Reward profiling: run swemer_v1's training tasks through the opencode sandboxed agent (not
+# the golden patch) to collect real rollouts + rewards for a policy checkpoint, 8 repeats per
+# task. See resources_servers/swemer_v1/configs/swemer_v1.yaml for the base resources server.
+config_paths:
+- responses_api_agents/opencode_sandboxed_agent/configs/opencode_sandboxed_agent.yaml
+- resources_servers/swemer_v1/configs/swemer_v1.yaml
+
+swemer_v1_reward_profiling_resources_server:
+  _inherit_from: swemer_v1_resources_server
+
+  resources_servers:
+    swemer_v1:
+      allowed_agents: [opencode_sandboxed_agent]
+      evaluation_timeout: 1200
+
+swemer_v1_reward_profiling_opencode_sandboxed_agent:
+  _inherit_from: opencode_sandboxed_agent
+  responses_api_agents:
+    opencode_sandboxed_agent:
+      resources_server:
+        type: resources_servers
+        name: swemer_v1_reward_profiling_resources_server
+      # 1hr harness execution budget per rollout, not the eval-server's own evaluation_timeout.
+      sandbox_timeout: 3600
+      # Keep every full turn in the rollout for reward profiling -- no OpenCode-side
+      # auto-summarization/pruning of earlier context.
+      opencode_config:
+        compaction:
+          auto: false
+          prune: false
+      datasets:
+        - name: swemer_v1_reward_profiling
+          type: benchmark
+          jsonl_fpath: resources_servers/swemer_v1/data/swemer_v1_training_enriched_with_prompt_template.jsonl
+          prepare_script: benchmarks/reward_profiling/swemer_v1/prepare.py
+          num_repeats: 8