feat(toolsandbox): add ToolSandbox multi-turn tool-use benchmark - #1949
Merged
ritaneves merged 18 commits intoJul 29, 2026
Conversation
marta-sd
reviewed
Jul 9, 2026
| # | ||
| # NOTE: scipy is intentionally NOT listed. Upstream used | ||
| # scipy.optimize.linear_sum_assignment for milestone scoring, but scipy is in | ||
| # nemo-gym's [tool.uv] exclude-dependencies and uv strips it from any install |
| # limitations under the License. | ||
| """Agent harness for the ToolSandbox resources server. | ||
|
|
||
| A thin fork of ``aviary_agent``. The difference: ToolSandbox is a |
Contributor
There was a problem hiding this comment.
Would it be possible to extend the aviary_agent instead? I think it's worth consolidating the two for future maintenance cost.
Collaborator
Author
There was a problem hiding this comment.
A fork here means mostly the structure. I've checked and their logic is way different and it's hard to find something to make it generic.
| gym env start \ | ||
| --config resources_servers/toolsandbox/configs/toolsandbox.yaml \ | ||
| --model-type openai_model \ | ||
| ++policy_base_url=https://openrouter.ai/api/v1 \ |
Contributor
There was a problem hiding this comment.
These overrides can be replaced with flags, see the mapping
Suggested change
| ++policy_base_url=https://openrouter.ai/api/v1 \ | |
| --model-url https://openrouter.ai/api/v1 \ |
| --config resources_servers/toolsandbox/configs/toolsandbox.yaml \ | ||
| --model-type openai_model \ | ||
| ++policy_base_url=https://openrouter.ai/api/v1 \ | ||
| ++policy_api_key=$OPEN_ROUTER_KEY \ |
Contributor
There was a problem hiding this comment.
Suggested change
| ++policy_api_key=$OPEN_ROUTER_KEY \ | |
| --model-api-key $OPEN_ROUTER_KEY \ |
| --model-type openai_model \ | ||
| ++policy_base_url=https://openrouter.ai/api/v1 \ | ||
| ++policy_api_key=$OPEN_ROUTER_KEY \ | ||
| ++policy_model_name=qwen/qwen3.5-9b |
Contributor
There was a problem hiding this comment.
Suggested change
| ++policy_model_name=qwen/qwen3.5-9b | |
| --model qwen/qwen3.5-9b |
| @@ -0,0 +1,320 @@ | |||
| # For licensing see accompanying LICENSE file. | |||
Contributor
There was a problem hiding this comment.
Consider removing this directory. Based on the info in VENDORING.md (L44) I conclude that this dir is not needed by the server. The maintenance cost might outweigh the benefits of having the ability to run offline checks.
| return kwargs | ||
|
|
||
|
|
||
| class OpenAIAPIAgent(BaseRole): |
Contributor
There was a problem hiding this comment.
If you decide to remove the cli dir, this class is no longer needed (I think)
Native gym port of apple/ToolSandbox — a stateful, conversational, tool-using benchmark — built on the aviary env pattern. The agent-under-test is the gym policy model driven by the toolsandbox_agent harness; the user simulator and the Python execution environment run inside the resources server; verify() is pure scoring. Reward is the milestone similarity in [0, 1] (0 if a minefield is hit). - resources_servers/toolsandbox — ToolSandboxResourcesServer with the aviary env lifecycle: seed_session -> obs + tools, /step(action) -> obs + done, /close (computes and caches the milestone/minefield breakdown), /verify (returns the cached reward). Each seed/step/close runs its role work inside its own asyncio.Task so concurrent episodes never share the vendored process-global ambient ExecutionContext. - responses_api_agents/toolsandbox_agent — thin fork of aviary_agent that forwards the whole model output (assistant text AND tool calls) to /step and lets the resources server decide done, so a natural-language turn routes to the user simulator instead of ending the rollout. - Vendored tool_sandbox/ (Apple custom source license) is used unchanged except for the conversation driver; scipy.optimize.linear_sum_assignment is replaced by a vendored dependency-free solver (scipy is excluded from the gym install), validated bit-identical to scipy. See tool_sandbox/VENDORING.md. - prepare_toolsandbox.py writes data/test.jsonl (one row per scenario index); committed example.jsonl is the 5-row smoke set. - toolsandbox.yaml wires the server + agent for local eval/training; toolsandbox_serve.yaml serves both for nemo-evaluator via the gym:// adapter. - Opt-in heavy deps live in requirements.txt (per-server .venv); ATTRIBUTIONS.md and pyproject ruff exclude cover the vendored tree. - Unit tests cover the scenario index, seed/step/close/verify, tool-call and natural-language routing, error observations, context isolation, the message cap, and user-sim extra_body flattening (19 tests). Signed-off-by: mbagdasarova <mbagdasarova@nvidia.com>
Signed-off-by: mbagdasarova <mbagdasarova@nvidia.com>
…iguration Signed-off-by: mbagdasarova <mbagdasarova@nvidia.com>
Signed-off-by: mbagdasarova <mbagdasarova@nvidia.com>
Signed-off-by: mbagdasarova <mbagdasarova@nvidia.com>
…uilders Signed-off-by: mbagdasarova <mbagdasarova@nvidia.com>
Signed-off-by: mbagdasarova <mbagdasarova@nvidia.com>
pachmu
commented
Jul 24, 2026
| | Component | Type | License | Author / Copyright | URL | | ||
| |------------|------|----------|--------------------|-----| | ||
| | **ToolSandbox** | Tool-use benchmark (scenarios, tools, sandbox, scoring) | Original: Apple MIT License (see full text below). NVIDIA modifications: Apache-2.0 | Copyright (C) 2024 Apple Inc.; Modifications Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES | https://github.com/apple/ToolSandbox | | ||
|
|
Collaborator
Author
There was a problem hiding this comment.
Vendored code approval: https://nvbugspro.nvidia.com/bug/5519198
ritaneves
self-requested a review
July 27, 2026 15:29
ritaneves
previously approved these changes
Jul 27, 2026
Contributor
|
/ok to test |
@ritaneves, there was an error processing your request: See the following link for more information: https://docs.gha-runners.nvidia.com/cpr/e/1/ |
Contributor
|
/ok to test 639483a |
Signed-off-by: mbagdasarova <mbagdasarova@nvidia.com>
…PDX license information Signed-off-by: mbagdasarova <mbagdasarova@nvidia.com>
… modifications Signed-off-by: mbagdasarova <mbagdasarova@nvidia.com>
…ications Signed-off-by: mbagdasarova <mbagdasarova@nvidia.com>
…nd tests Signed-off-by: mbagdasarova <mbagdasarova@nvidia.com>
Signed-off-by: mbagdasarova <mbagdasarova@nvidia.com>
pachmu
force-pushed
the
mbagdasarova/toolsandbox-integr
branch
from
July 27, 2026 16:05
639483a to
41d17f8
Compare
ritaneves
self-requested a review
July 27, 2026 16:14
ritaneves
approved these changes
Jul 27, 2026
ko3n1g
approved these changes
Jul 29, 2026
OlegSudakov
pushed a commit
to OlegSudakov/Gym
that referenced
this pull request
Aug 7, 2026
…DIA-NeMo#1949) Native gym port of apple/ToolSandbox — a stateful, conversational, tool-using benchmark — built on the aviary env pattern. The agent-under-test is the gym policy model driven by the toolsandbox_agent harness; the user simulator and the Python execution environment run inside the resources server; verify() is pure scoring. Reward is the milestone similarity in [0, 1] (0 if a minefield is hit). - resources_servers/toolsandbox — ToolSandboxResourcesServer with the aviary env lifecycle: seed_session -> obs + tools, /step(action) -> obs + done, /close (computes and caches the milestone/minefield breakdown), /verify (returns the cached reward). Each seed/step/close runs its role work inside its own asyncio.Task so concurrent episodes never share the vendored process-global ambient ExecutionContext. - responses_api_agents/toolsandbox_agent — thin fork of aviary_agent that forwards the whole model output (assistant text AND tool calls) to /step and lets the resources server decide done, so a natural-language turn routes to the user simulator instead of ending the rollout. - Vendored tool_sandbox/ (Apple custom source license) is used unchanged except for the conversation driver; scipy.optimize.linear_sum_assignment is replaced by a vendored dependency-free solver (scipy is excluded from the gym install), validated bit-identical to scipy. See tool_sandbox/VENDORING.md. - prepare_toolsandbox.py writes data/test.jsonl (one row per scenario index); committed example.jsonl is the 5-row smoke set. - toolsandbox.yaml wires the server + agent for local eval/training; toolsandbox_serve.yaml serves both for nemo-evaluator via the gym:// adapter. - Opt-in heavy deps live in requirements.txt (per-server .venv); ATTRIBUTIONS.md and pyproject ruff exclude cover the vendored tree. - Unit tests cover the scenario index, seed/step/close/verify, tool-call and natural-language routing, error observations, context isolation, the message cap, and user-sim extra_body flattening (19 tests). --------- Signed-off-by: mbagdasarova <mbagdasarova@nvidia.com>
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
4 participants
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.
Native gym port of apple/ToolSandbox — a stateful, conversational, tool-using benchmark — built on the aviary env pattern. The agent-under-test is the gym policy model driven by the toolsandbox_agent harness; the user simulator and the Python execution environment run inside the resources server; verify() is pure scoring. Reward is the milestone similarity in [0, 1] (0 if a minefield is hit).