Repository navigation
feat(swebench-pro): configure Hermes for SWE-bench Pro - #3722
Merged
yaoyu-33 merged 1 commit intoSep 30, 2026
Conversation
ananthsub
force-pushed
the
ananthsub/episode-orchestration-prototype
branch
from
September 28, 2026 15:41
e956be3 to
c6a59f2
Compare
yaoyu-33
added
area:env-infra
ananthsub
force-pushed
the
ananthsub/episode-orchestration-prototype
branch
from
September 28, 2026 15:48
c6a59f2 to
4ccda82
Compare
ananthsub
force-pushed
the
ananthsub/episode-orchestration-prototype
branch
from
September 28, 2026 16:03
4ccda82 to
d92a1e3
Compare
ananthsub
force-pushed
the
ananthsub/episode-orchestration-prototype
branch
from
September 28, 2026 19:43
4878020 to
4834972
Compare
ananthsub
force-pushed
the
ananthsub/episode-orchestration-prototype
branch
from
September 28, 2026 19:54
4834972 to
41565c3
Compare
ananthsub
force-pushed
the
ananthsub/episode-orchestration-prototype
branch
from
September 28, 2026 20:12
41565c3 to
693ba6f
Compare
ananthsub
force-pushed
the
ananthsub/episode-orchestration-prototype
branch
from
September 28, 2026 20:38
693ba6f to
09cc1a9
Compare
This was referenced Sep 29, 2026
Adds benchmarks/swebench/pro/hermes.yaml, which runs SWE-bench Pro rows through single_agent_turn_legacy with Hermes executing in the task sandbox, alongside the existing opencode.yaml. The configuration composes the single swebench_pro.yaml Resources Server config, which now lists hermes_agent among its allowed agents. Signed-off-by: Ananth Subramaniam <ansubramania@nvidia.com>
Contributor
|
🔄 Merge queue validation started! You can track the progress here: https://github.com/NVIDIA-NeMo/Gym/actions/runs/36650451687 |
This branch was previously deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What does this PR do?
Adds the configuration that runs SWE-bench Pro with Hermes through an Environment Server:
benchmarks/swebench/pro/hermes.yaml, alongside the existingopencode.yaml. It runs SWE-bench Pro rows throughsingle_agent_turn_legacy, with the SWE-bench Pro Resources Server preparing the task sandbox and Hermes working inside it.resources_servers/swebench_pro/configs/swebench_pro.yamllistshermes_agentamong its allowed agents, so the Hermes configuration composes the same Resources Server config as the other SWE-bench Pro agents.Rows reach the Environment Server by their agent, as with
opencode.yaml: one Environment Server fronts the Hermes agent in this configuration.Simple Agent sessions and the example Resources Server's sessions moved to their own PR, #3784.
Stack position
Part 3 of 3. Previous: #3721. This is the top of the stack.
No issue is linked because this is one layer of the Environment Server stack.
Validation
gym env resolve --config benchmarks/swebench/pro/hermes.yaml ...resolves to the Hermes agent, the SWE-bench Pro Resources Server, and thesingle_agent_turn_legacyEnvironment Server, with that Environment Server as the only one fronting the Hermes agent.pytest tests/unit_tests/test_task_data.py tests/unit_tests/test_global_config.py tests/unit_tests/test_benchmarks.py -q: 572 passed.pre-commit runon the changed files.Rollouts
Real Hermes + SWE-bench Pro rollouts with
gpt-5.6-solran throughsingle_agent_turn_legacyon this stack, with Hermes calling the Model Server directly from the task sandbox. The task resolved with reward1.0(300 of 300 tests) in two runs and passed 299 of 300 tests in two others, with empty failure sidecars and no leftover sandboxes. Those runs selected the Environment Server explicitly; this configuration reaches the same Environment Server by agent routing.Compatibility and benchmark impact
Configuration only. Existing SWE-bench Pro configurations are unchanged apart from the added allowed agent.