Repository navigation
docs: shorten benchmark onboarding guidance - #3713
Conversation
|
🌿 Preview your docs: https://nvidia-preview-rfneves-gym-benchmark-onboarding-docs.docs.buildwithfern.com/nemo/gym Here are the markdown pages you've updated: |
|
/claude review |
|
SHIP — LGTM, no reliability concerns. Docs-only change to two Verified against the trusted base source:
No findings. |
5b77680 to
8cc96d9
Compare
Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>
yaoyu-33
left a comment
There was a problem hiding this comment.
Fixed the outdated architecture wording in 6eb93be: Environment Server routing is already required; legacy forwarding and native protocol migration are now distinguished. Docs/link checks and pre-commit pass (authenticated redirect checking unavailable). LGTM.
|
/ok to test 6eb93be |
|
🔄 Merge queue validation started! You can track the progress here: https://github.com/NVIDIA-NeMo/Gym/actions/runs/37030616925 |
|
🔄 Merge queue validation started! You can track the progress here: https://github.com/NVIDIA-NeMo/Gym/actions/runs/37031058893 |
| | `data/source.jsonl`, `data/example.jsonl` | Replace the sample task and retain representative examples | | ||
| | `README.md` | Measurement contract, preparation and evaluation commands, and limitations | | ||
|
|
||
| It also creates `resources_servers/my_benchmark/` with a verifier implementation, config, dependencies, unit tests, and verifier fixture. Reuse a harness for a new task; a new dataset or prompt does not by itself require a new Agent Server. See [Single Step Environment](/main/environment-tutorials/single-step-environment) when implementing new task tools or scoring. |
There was a problem hiding this comment.
We'll need to make a note to update this as we are defining the task definition
| The most important principle when adding benchmarks into NeMo Gym is ensuring the fidelity of the benchmark. As a result, there are additional steps and best practices to adding a benchmark that are required on top of adding just a training environment (although the steps below are still suggested for training environments). | ||
|
|
||
| ## Reward profiling | ||
| # Add a benchmark |
There was a problem hiding this comment.
It might be good to have a "hello world" example. The detail below is excellent. But it might be good to show an overview so users are not intimidated by the full detail, and they percieve the onboarding to be straight forward.
This can be a follow up, but do think its important.
Flow would be
- Overview of steps 1-5
- Hello world basic example
- Step 1
- ...
- Step 5
…IA-NeMo#4043) ## What does this PR do? Follow up on the review of NVIDIA-NeMo#3713 by introducing benchmark onboarding with a five-step overview and a concrete hello-world example before the detailed instructions. The example uses the generated one-question scaffold, explains its prompt and exact-match reward, and shows an optional one-row model evaluation. Make local publication an explicit fifth step. Add task-contract maintenance guidance and a maintainer reminder to update the scaffold file/responsibility mapping as the shared task definition evolves. ## Validation - `pre-commit run --all-files` — passed. - `python tests/unit_tests/test_fern_docs_links.py` — 18 tests and 17 subtests passed. - `fern check --warnings` — 0 errors; authenticated redirect checking skipped because Fern login is unavailable. - Generated scaffold checked through the Python API in a temporary directory: preparation output, prompt rendering without the reference answer, manifest parsing, and correct/incorrect/malformed verifier fixture cases. - `git diff --check` — passed. Full CLI execution was not verified because the existing local virtual environment lacks `httptools`. No model evaluation was run. ## Compatibility and benchmark impact Documentation only; no runtime, verifier, or benchmark-result changes. Frozen release documentation is unchanged. The commit includes DCO sign-off. --------- Signed-off-by: Rita Fernandes Neves <rfernandesne@nvidia.com> Signed-off-by: ritaneves <rfernandesne@nvidia.com>
What does this PR do?
Shorten benchmark onboarding into scaffold, prepare, validate/test, smoke, and local publish steps. Clarify manifest prompting and verifier reuse, and distinguish current Environment Server/legacy adapter routing from native episode-protocol migration (#1866, #3724).
Validation
Validated at
6eb93be01:pre-commit run --all-files— passed.python tests/unit_tests/test_fern_docs_links.py— 18 passed.python -m pytest --import-mode=importlib tests/unit_tests/test_global_config.py -k config_without_environment_server_is_rejected -o addopts='' -q— 1 passed.cd fern && npm run check -- --warnings— 0 errors; authenticated redirects check skipped because Fern login is unavailable.git diff --check— passed.Rollouts
N/A: documentation-only changes; no runtime, agent, or verifier behavior changes.
Compatibility and benchmark impact
No runtime or benchmark-result changes. The guide now describes the existing requirement that runnable agents have Environment Server routing; compatibility routing is distinct from migration to native protocols.
Checklist
pre-commit run --all-files) (so CI lint/format checks pass).git commit -s) (so the DCO check passes).