docs(fern): rewrite key terminology glossary (#1500) by sephmard · Pull Request #2355 · NVIDIA-NeMo/Gym

sephmard · 2026-08-05T18:14:39Z

Summary

Rewrite Key Terminology with Gym-aligned definitions (Environment, Agent, Rollout, Benchmark, etc.)
Add an Overview (run flow, Agent/Environment composition, Model server) plus a concept → Gym component map
Cross-link related how-tos (Data, Agent/Model Server, Verifiers, Evaluate, Training, Sandboxes)

Test plan

make docs-check (0 errors)
python3 tests/unit_tests/test_fern_docs_links.py
Confirm Fern docs preview comment on the PR
Skim Overview + glossary sections for clarity on the preview URL

Align definitions with Environments/Evaluation/Training/Architecture, lead with a Gym overview and component map, and link out to how-to pages. Signed-off-by: Seph Mard <smard@nvidia.com>

copy-pr-bot · 2026-08-05T18:14:43Z

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

github-actions · 2026-08-05T18:15:42Z

🌿 Preview your docs: https://nvidia-preview-docs-update-key-terminology-1500.docs.buildwithfern.com/nemo/gym

Here are the markdown pages you've updated:

ritaneves · 2026-08-05T18:20:29Z

+| **Environment** | **Dataset** + **Agent harness** + **Verifier** + **State** | Everything needed to run and score a task. The **model is outside** the environment. |

-**Task Instance**
+The **model** is the LLM you call — local weights or a remote API endpoint. In Gym it is usually exposed by the **Model server** (or your harness can call an endpoint directly). The same environment can be used as a **benchmark** (fixed eval) or a **training environment** (learn from rewards / synthetic data).


nit: may be a vlm, cf super

Went with modality-neutral wording instead of listing LLM/VLM: model / policy is what you call for generation (local weights or remote endpoint). That covers text, vision-language, and other modalities without baking a modality into the glossary.

Co-authored-by: Felipe Vieira Frujeri <ffrujeri@nvidia.com> Signed-off-by: Seph Mard <seph.mard@gmail.com>

ritaneves · 2026-08-05T18:27:08Z

+| **SFT** | Supervised fine-tuning from examples of good behavior (demonstration data). |
+| **RL** | Improve the policy through environment interaction and reward signals. |
+| **Online / offline** | **Online**: update the policy from rewards while interacting (e.g. GRPO). **Offline**: train from pre-collected rollouts (e.g. SFT, DPO). See [Offline training with rollouts](/tutorials/training-tutorials/offline-training-w-rollouts). |
+| **DPO** | Offline preference training from pairs of rollouts. |


nit: i would expand the acronyms

Addressed in 5cd0537 / earlier review commit: expanded SFT (Supervised Fine-Tuning), RL (Reinforcement Learning), DPO (Direct Preference Optimization), and GRPO (Group Relative Policy Optimization).

Addressed — full names are expanded in the Training glossary.

ritaneves · 2026-08-05T18:29:08Z

+| --- | --- |
+| Dataset | JSONL (`responses_create_params` + `verifier_metadata`) — [Prepare Data](/data) |
+| Agent harness | **Agent server** (`responses_api_agents/`) — [Agent Server](/agent-server) |
+| Verifier + state (+ env tools) | **Resources server** (`resources_servers/`) — [Build Verifiers](/build-verifiers) |


this is confusing to me. the state is obtained by a snapshot of the environment, no?

Clarified: per-attempt state is the mutable session world inside the resources server (seed_session → tools mutate it → verify()), not a snapshot of the whole environment config. Mapping table updated accordingly.

ritaneves · 2026-08-05T18:30:00Z

-A single rollout attempt for a specific task. Multiple instances per task capture different approaches.
+| Concept | NeMo Gym component |
+| --- | --- |
+| Dataset | JSONL (`responses_create_params` + `verifier_metadata`) — [Prepare Data](/data) |


I think calling this a component is confusing. I would put the link to the data section, and the function names

Addressed: retitled the map to “In NeMo Gym” and described Dataset as JSONL row fields + Prepare Data (not as a server “component”).

ffrujeri · 2026-08-05T18:28:57Z

+| | Built from | Notes |
+| --- | --- | --- |
+| **Agent** | **Model** + **Agent harness** | The harness turns model calls into a loop with tools and stop conditions. |
+| **Environment** | **Dataset** + **Agent harness** + **Verifier** + **State** | Everything needed to run and score a task. The **model is outside** the environment. |


One important distinction here is that the Agent harness is not a responsibility of the Environment
it is part of the Agent as you described in the line above. But the Environment packaging that we are doing contains a reference to the Agent Server

| Environment | Dataset + Agent harness + Verifier + State | Everything needed to run and score a task. The model is outside the environment. |

Addressed: Environment is now Dataset + Verifier + State (+ env tools). Explicit note that the harness belongs to the Agent; Gym configs may still reference an agent server for packaging only.

ffrujeri · 2026-08-05T18:38:15Z

-**Data Generation Process**
+| Term | Definition |
+| --- | --- |
+| **Environment** | Everything required for an agent to complete a task except the model: `Dataset + Agent harness + Verifier + State`. See [Environments](/about/concepts/environments) and [Build Environments](/environment-tutorials). |


Same recommendation as before.

Addressed: Environment glossary row matches Overview — harness is a packaging reference, not part of the Environment formula.

ffrujeri · 2026-08-05T18:39:46Z

-**Preference Pairs**
+| Term | Definition |
+| --- | --- |
+| **Training environment** | Environment used to improve a model — as an online RL reward signal, or to generate synthetic training data. See [Training](/about/concepts/training). |


not sure if there is need to make a distinction between a "Training environment" from an Environment apart from the taskset and data contamination

Agreed — dropped “Training environment” as a separate type. Overview now says the same environment shape is used for benchmark eval vs training; differences are mainly task split / contamination controls.

Co-authored-by: Felipe Vieira Frujeri <ffrujeri@nvidia.com> Signed-off-by: Seph Mard <seph.mard@gmail.com>

Clarify agent vs environment ownership, expand training acronyms, treat models as LLM/VLM, and reword the Gym mapping table. Signed-off-by: Seph Mard <smard@nvidia.com>

Generalize model definitions so they are not LLM-specific. Signed-off-by: Seph Mard <smard@nvidia.com>

…IA-NeMo#2355) ## Summary - Rewrite [Key Terminology](https://docs.nvidia.com/nemo/gym/main/about/concepts/key-terminology/) with Gym-aligned definitions (Environment, Agent, Rollout, Benchmark, etc.) - Add an Overview (run flow, Agent/Environment composition, Model server) plus a concept → Gym component map - Cross-link related how-tos (Data, Agent/Model Server, Verifiers, Evaluate, Training, Sandboxes) Fixes NVIDIA-NeMo#1500 ## Test plan - [x] `make docs-check` (0 errors) - [x] `python3 tests/unit_tests/test_fern_docs_links.py` - [x] Confirm Fern docs preview comment on the PR - [x] Skim Overview + glossary sections for clarity on the preview URL --------- Signed-off-by: Seph Mard <smard@nvidia.com> Signed-off-by: Seph Mard <seph.mard@gmail.com> Co-authored-by: Felipe Vieira Frujeri <ffrujeri@nvidia.com>

docs(fern): rewrite key terminology glossary (#1500)

08cac1c

Align definitions with Environments/Evaluation/Training/Architecture, lead with a Gym overview and component map, and link out to how-to pages. Signed-off-by: Seph Mard <smard@nvidia.com>

Merge branch 'main' into docs/update-key-terminology-1500

7e23125

sephmard requested review from ananthsub, ffrujeri and lbliii August 5, 2026 18:15

sephmard added the documentation Improvements to documentation label Aug 5, 2026

sephmard self-assigned this Aug 5, 2026

ritaneves reviewed Aug 5, 2026

View reviewed changes

ffrujeri reviewed Aug 5, 2026

View reviewed changes

Comment thread fern/versions/latest/pages/about/concepts/index.mdx Outdated

Update fern/versions/latest/pages/about/concepts/index.mdx

79c1e50

Co-authored-by: Felipe Vieira Frujeri <ffrujeri@nvidia.com> Signed-off-by: Seph Mard <seph.mard@gmail.com>

ritaneves reviewed Aug 5, 2026

View reviewed changes

bxyu-nvidia previously approved these changes Aug 5, 2026

View reviewed changes

ffrujeri reviewed Aug 5, 2026

View reviewed changes

Update fern/versions/latest/pages/about/concepts/key-terminology.mdx

8df9b0d

Co-authored-by: Felipe Vieira Frujeri <ffrujeri@nvidia.com> Signed-off-by: Seph Mard <seph.mard@gmail.com>

sephmard dismissed bxyu-nvidia’s stale review via 8df9b0d August 5, 2026 18:43

sephmard and others added 4 commits August 5, 2026 14:43

Update fern/versions/latest/pages/about/concepts/key-terminology.mdx

54d57ac

Co-authored-by: Felipe Vieira Frujeri <ffrujeri@nvidia.com> Signed-off-by: Seph Mard <seph.mard@gmail.com>

Update fern/versions/latest/pages/about/concepts/key-terminology.mdx

6e1c237

Co-authored-by: Felipe Vieira Frujeri <ffrujeri@nvidia.com> Signed-off-by: Seph Mard <seph.mard@gmail.com>

docs(fern): address key terminology review feedback

2fd8e42

Clarify agent vs environment ownership, expand training acronyms, treat models as LLM/VLM, and reword the Gym mapping table. Signed-off-by: Seph Mard <smard@nvidia.com>

docs(fern): use modality-neutral model/policy wording

5cd0537

Generalize model definitions so they are not LLM-specific. Signed-off-by: Seph Mard <smard@nvidia.com>

ffrujeri self-requested a review August 5, 2026 19:54

ffrujeri approved these changes Aug 5, 2026

View reviewed changes

Merge branch 'main' into docs/update-key-terminology-1500

2478228

sephmard merged commit 5da2f51 into main Aug 5, 2026
10 checks passed

sephmard deleted the docs/update-key-terminology-1500 branch August 5, 2026 20:48

Uh oh!

Conversation

sephmard commented Aug 5, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Test plan

Uh oh!

copy-pr-bot Bot commented Aug 5, 2026

Uh oh!

github-actions Bot commented Aug 5, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

sephmard commented Aug 5, 2026 •

edited

Loading