feat: document --agent-type CLI flag and agent swapping mechanism - #2793
Conversation
|
🌿 Preview your docs: https://nvidia-preview-martas-agent-selector-docs.docs.buildwithfern.com/nemo/gym Here are the markdown pages you've updated: |
sephmard
left a comment
There was a problem hiding this comment.
Approving. I also think that we need to include an update in the gym/evaluation tab. I'd suggest that this is its own new subtab under Evaluate: Harness.
The merge-base changed after approval.
--agent-type CLI flag and agent swapping mechanism
9ffc9b8 to
5e8ecfc
Compare
|
Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually. Contributors can view more details about this message here. |
3a4c2b3 to
b301c4c
Compare
sephmard
left a comment
There was a problem hiding this comment.
This page is about agentic evaluation, but "The composed instance is renamed" example (gpqa_mcqa_simple_agent/hermes_agent) isn't agentic — gpqa / gsm8k uses the mcqa resources server, which is plain multiple-choice verification, no tools.
Please swap it for working agentic example — the intro's own terminal_bench_2_1/opencode + terminus_2_agent composition works directly here.
| # Run GPQA with the Hermes harness instead of the one its config names | ||
| gym eval run --agent-type hermes_agent --benchmark gpqa --model-type vllm_model |
There was a problem hiding this comment.
GPQA isnt the best example, as its non-agentic
| ## The composed instance is renamed | ||
|
|
||
| Composition replaces the environment's agent and renames the server instance after the harness that | ||
| runs it, so metrics and `gym env status` report what actually ran. Composing `hermes_agent` onto | ||
| `gpqa_mcqa_simple_agent` produces `gpqa_mcqa_hermes_agent`. | ||
|
|
||
| That name is what you use to route rollouts when the servers are already up: | ||
|
|
||
| ```bash | ||
| gym env start --agent-type hermes_agent --benchmark gpqa --model-type vllm_model | ||
| gym eval run --no-serve --agent gpqa_mcqa_hermes_agent --input rows.jsonl | ||
| ``` | ||
|
|
||
| End-to-end rollouts collection (without the `--no-serve` flag) keeps working — rows stamped with the | ||
| pre-swap instance are re-routed for you under the hood. |
There was a problem hiding this comment.
"The composed instance is renamed" example (gpqa_mcqa_simple_agent/hermes_agent) isn't agentic (plain MCQA, no tools); suggest swapping to the intro's own terminal_bench_2_1/opencode + terminus_2_agent composition, both for accuracy-to-theme and to cut the page down to one running example instead of three.
There was a problem hiding this comment.
Replacing it requires introducing changes to the configs - otherwise the composed name was completely unreadable. This change requires approval from eng team (ideally @bxyu-nvidia) because it touches configs that are in the critical path.
There was a problem hiding this comment.
Ok, thanks for the context
Due to what's in configs the name would be "opencode_sandboxed_agent_terminal_bench_2_1_terminus_2_sandboxed_agent". I need to update configs for this to make sense. |
Signed-off-by: Marta Stepniewska-Dziubinska <martas@nvidia.com>
Signed-off-by: Marta Stepniewska-Dziubinska <martas@nvidia.com>
Signed-off-by: Marta Stepniewska-Dziubinska <martas@nvidia.com>
Signed-off-by: Marta Stepniewska-Dziubinska <martas@nvidia.com>
Co-authored-by: Seph Mard <seph.mard@gmail.com> Signed-off-by: Marta Stepniewska-Dziubinska <marta-sd@users.noreply.github.com>
Signed-off-by: Marta Stepniewska-Dziubinska <martas@nvidia.com>
Co-authored-by: Seph Mard <seph.mard@gmail.com> Signed-off-by: Marta Stepniewska-Dziubinska <marta-sd@users.noreply.github.com>
Server instances are named <benchmark>_<implementation>, but the sandbox benchmarks led with the harness. That also broke the composed instance name: the rename strips the agent type only as a suffix, so swapping a harness in appended the new type instead of replacing the old one. Signed-off-by: Marta Stepniewska-Dziubinska <martas@nvidia.com>
The rename and pairing examples used gpqa/gsm8k, which score multiple choice with no tools, on a page about agentic evaluation. All the sample output is now copied from real runs of the intro's own swap. Signed-off-by: Marta Stepniewska-Dziubinska <martas@nvidia.com>
ee41d5d to
46f5ce8
Compare
sephmard
left a comment
There was a problem hiding this comment.
approving docs changes - @bxyu-nvidia reviewed the config changes
This PR documents the new
--agentCLI flag and agent swapping mechanism. It should be merge as the last as it documents features that are not yet in themainbranch.The entire PR stack will introduce the following changes:
--agent-typeflag in the CLI for selecting the agent