> For clean Markdown of any page, append .md to the page URL. > For a complete documentation index, see https://docs.nvidia.com/nemo/gym/llms.txt. > For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.nvidia.com/nemo/gym/_mcp/server. # Release Notes #### v0.5.0 ### Release Summary The NeMo Gym 0.5.0 release expands the sandbox ecosystem to seven providers, adds four new general-purpose agent harnesses (Codex, KiloCode, RemoteAgent, and Any-SWE) bringing the total to 20, adds 21 new benchmarks and environments, and wires rollout observability end-to-end from the model server boundary through agent transcripts. Highlights: * Seven sandbox providers: Docker, Daytona, ECS Fargate, Enroot, and OpenShell join OpenSandbox and Apptainer; large-scale OpenSandbox reliability significantly improved * Four new agent harnesses: Codex CLI, KiloCode, RemoteAgent, and `anyswe_agent` * Recompute rewards from stored rollouts without re-running inference with `gym eval reverify` * Rollout observability joined end-to-end: model-call capture, agent observations, and a standardized `ng_trajectory` schema * 21 new environments across six domains: Agentic, Knowledge and instruction following, Long context, Science and coding, Translation and multilingual, and Reasoning ### First-Time Contributors We welcomed 22 new contributors to NeMo Gym with this release: * **@mpatel31415** added `gym eval reverify` command and `--judge-failed-only` flag for recovering failed judge rows * **@Glorf** added per-rollout model-call capture, Docker and ECS Fargate sandbox providers, rollout observation contract, Claude Code rollout observations, and standardized `ng_trajectory` schema * **@nblintao** added per-request policy endpoint override for the SWE agents, enabling RL training frameworks to route each episode through a per-episode recording proxy * **@JeffPengCoder** brought OSWorld — a stateful desktop GUI benchmark — into the benchmark catalog * **@rystewart-nvidia** added the Legal Agent Bench integration, exposing Harvey's 1,749-task LAB benchmark through standard Gym eval commands * **@fallintoplace** fixed the long-standing disagreement between runtime aggregate metrics and the persisted `output.jsonl` * **@jonathanlli** added RULER pretrain evaluation, enabling text-completion scoring for base and midtraining checkpoints * **@thompsonb** overhauled WMT24++ and FLORES translation evaluation (55 locales, 219 language pairs, chrF/spBLEU scoring); expanded MMLU-ProX to 29 languages * **@hkumar92** added the PinchBench agentic benchmark * **@pachmu** added ToolSandbox, IHEval, RoleMRC, and RAGTruth benchmarks Thank you to all 57 NeMo Gym contributors this cycle, including 22 first-time contributors! ### Command Line Interface * **`gym eval reverify`** — re-run only the verifier on stored rollouts; `--judge-failed-only` recovers rows that failed due to a flaky judge without re-verifying successful rollouts * `gym list` and `gym search` extended to cover models, resources-servers, and agents; `gym list ` drills into a single artifact * External plugins discoverable via `--search-dir` and environment variables; `-v/--verbose` now accepted before any subcommand ### Sandboxing Five new built-in providers: **Docker**, **Daytona**, **ECS Fargate**, **Enroot**, and **OpenShell** join OpenSandbox and Apptainer. Provider choice is a one-line config swap — any agent built on `nemo_gym.sandbox` works with any provider unchanged. OpenSandbox reliability at scale is significantly improved: keepalive-bounded transport eliminates silent rollout zeroing at concurrency 300–1500; image registry auth supports private container images; sandbox resources are automatically labeled with team, user, and workload identifiers. See [Available Sandbox Providers](/nemo-gym/nemo_gym/sandbox/providers) for the full list. ### Configure Agent Harnesses New harnesses join the existing set (Claude Code, Hermes, mini-SWE-Agent, OpenClaw, Pi, and more): * **Codex** and **KiloCode** integrate `codex exec` and `kilo run` respectively, routing model calls through Gym for per-rollout capture * **`anyswe_agent`** runs any Gym harness inside a SWE task container * **`swe_agents`** adds OpenCode as a supported agent framework alongside OpenHands, with DeepSWE and DeNovoSWE dataset support and message replay for trajectory branching * **RemoteAgent** drives any external service that implements `POST /v1/responses`, with Gym owning the tool loop and verification ### Configure Models * vLLM can now drive `/v1/completions` for base and pretrain checkpoint evaluation via opt-in `use_completions_api` * All Gym model servers now accept `stream: true` on `/v1/chat/completions` via synthesized SSE, unblocking streaming-first clients such as OpenClaw and Codex * Add `expose_tools_over_mcp: true` to any resources server config to serve its tools over MCP with no handler code changes ### Rollout Observability * Per-rollout model-call capture records requests, responses, token usage, and latency at the model server boundary * Claude Code transcripts populate `ng_agent_observations`; agent observations and model-call capture are joined through a standardized `ng_trajectory` schema * Judge failures are routed to a `_failures.jsonl` sidecar, keeping aggregate metrics over successfully-judged rows only ### New Benchmarks and Environments 21 new environments across six domains: * **Agentic:** PinchBench (147 real-world tasks), OSWorld (desktop GUI with VM-backed evaluation), Legal Agent Bench (1,749 Harvey LAB tasks), ToolSandbox (Apple multi-turn tool-use), BrowseComp (web research), BioMNIBench DA, Tau3 banking (BM25+grep offline eval path) * **Knowledge and instruction following:** SECQUE, FinanceBench, Finance SEC Search, IHEval (instruction hierarchy, rule-based), Litmus-Bench v0.1, RoleMRC (role-play MRC), RAGTruth (hallucination detection) * **Long context:** NIAH (retrieval with overlap penalty) * **Science and coding:** CVDP Agentic (expanded to support the agentic subset, harness-agnostic) * **Translation and multilingual:** WMT24++ (expanded from 5 to 55 locales), FLORES (expanded from 30 to 219 language pairs, chrF/spBLEU scoring), MMLU-ProX (expanded to 29 languages); RULER now supports pretrain text-completion evaluation * **Reasoning:** ReasoningGym environments — six agentic variants: Claude Code, Hermes, and four LangGraph-based variants (orchestrator, reflection, parallel thinking, and ReWOO) See the [Available Environments](https://github.com/NVIDIA-NeMo/Gym#-available-environments) table for the full list. ### Deprecation Notices * WMT24++ and FLORES scores from prior versions are not comparable with this version's chrF/spBLEU output * Python 3.13.14 is now required (previously 3.12); users running Gym in Python 3.12 environments must upgrade ### Bug Fixes * SciCode realigned to the AA 65-problem test set with per-rollout subtask accuracy reporting * Fixed silent rollout zeros at high concurrency against OpenSandbox (keepalive-bounded transport) * Fixed frozen rollouts in long-running benchmarks (TCP keepalive on global aiohttp connector) * Fixed `gym eval run` failing with `FileNotFoundError` when the output directory did not exist * Fixed `tool_choice` sent to vLLM without `tools`, causing request rejection * Fixed Claude Code `max_turns` hardcoded to 30; `max_turns: null` now removes the cap * Fixed Apptainer sandbox env vars injected into subprocess argv instead of environment * Fixed MCQA answer parsing for wrapped formats (`$D$`, `(D)`, `\boxed{\text{Answer: G}}`) * Fixed aggregate metrics including non-persisted rollouts, causing disagreement with `output.jsonl` ### Documentation * Rewrote the key terminology glossary with a Gym overview, component map, and links to how-to pages * New page documenting the Anthropic Messages dialect (`POST /v1/messages`) and wiring Claude Code through a Gym model server * Updated NeMo RL v0.7.0 compatibility guidance * Migrated all remaining docs examples from legacy `ng_run`/`ng_collect_rollouts` to the unified `gym` CLI * Documented `gym list` and `gym search` extensions, external plugin discovery, and MCP auto-exposure * Added `gym eval reverify` and multi-reward verification contract documentation ### Release Assets [GitHub Release v0.5.0](https://github.com/NVIDIA-NeMo/Gym/releases/tag/v0.5.0) #### v0.4.0 ### Release Summary NeMo Gym v0.4.0 expands evaluation tooling and agent integrations. It establishes a new monthly release cadence; we will continue to provide day-zero support for Nemotron models, datasets, and environments. Highlights: * Unified `gym` CLI: find agents and benchmarks by name with `gym list`, and catch config mistakes early with `gym env validate` * Diagnose evaluations with BLADE, an analysis skill for agents that reads your evaluation results and produces an evidence-backed report of which tasks failed, why, and the highest-impact fix (e.g. to the agent harness, training, verifier, or prompt) * Measure the impact of agent skills: run the same tasks with different skill sets and compare how each changes agent performance * Run agents in isolated sandboxes through a new pluggable provider framework * More agent harnesses out of the box, including OpenClaw, Pi, and OpenCode * Connect to hosted inference providers: Fireworks, Together.ai, OpenRouter, and more * New benchmarks across science, long-context, and interactive tasks ### First-Time Contributors We welcomed 20+ new contributors to this release! A few highlights: * **@marta-sd** and **@wprazuch** led the CLI refactor and clearer config errors * **@hemildesai** added the pluggable sandbox provider infrastructure and OpenSandbox as the first built-in * **@adil-a** laid the groundwork for Gym-owned MCP resources servers, letting a server expose its tools over MCP * **@eric-tramel** added the BunsenChem chemistry benchmark * **@jeffwillette** added the long machine translation datasets and servers Thank you to all the new contributors for helping make NeMo Gym better! ### Command Line Interface * One `gym` command for the full workflow, with `gym env`, `gym eval`, `gym list`, and `gym dataset` subcommands * Reference agents, benchmarks, and environments by name: use `gym list` to see what is available * `gym env validate` checks your config for missing, malformed, or empty values before a run and reports actionable errors ### Evaluation & Diagnostics * Skill evaluation: measure how agent skills affect performance by running the same tasks with different skill sets. Skills apply at rollout time as a run-level knob, so one dataset works across all skill variants and every rollout is tagged for comparison * BLADE (Benchmark Level Analysis and Diagnostics Engine): a built-in analysis skill that reads an agent run's rollouts, metrics, and configs and produces an evidence-backed report of which tasks failed, why, and the highest-impact fix (e.g. harness, training, verifier, or prompt) ### Sandboxing * Run tool-using and coding agents in isolated sandboxes through a pluggable provider framework * Built-in OpenSandbox and Apptainer providers, with third-party providers discoverable via entry points ### Configure Agent Harnesses New harnesses join the existing built-in set (Claude Code, Hermes, OpenHands, and more): * Added OpenCode, OpenClaw, and Pi agents for evaluation * Claude Code runtime capabilities (tool access, MCP servers, and bare vs. native auto-discovery mode) are now easily set via the server config ### Configure Models * New `inference_provider` model server connects to any OpenAI-compatible hosted provider (Fireworks, Together.ai, OpenRouter, DeepInfra, Gemini, and more) with ready-made configs * Every Gym model server now speaks the Anthropic Messages API, so Anthropic-native harnesses like the Claude Code CLI can run against any model you serve with Gym — see [Anthropic Messages](/model-server/anthropic-messages) ### New Benchmarks * **Science:** CritPt (research-level physics), SciCode (scientific coding), BunsenChem (chemistry multiple-choice), and FrontierScience Research (rubric-scored science) * **Long context:** Graphwalks (long-context graph reasoning) and Long Machine Translation (PG19, WMT24++) * **Interactive:** TALES, a text-adventure game suite See the [Available Environments](https://github.com/NVIDIA-NeMo/Gym#-available-environments) table for the full list. ### Deprecation Notices * The legacy `ng_*` and `nemo_gym_*` CLI commands (such as `ng_run` and `ng_collect_rollouts`) are deprecated in favor of the unified `gym` CLI. They still work for now but will be removed in a future release. ### Bug Fixes * Fixed intermittent connection errors during high-concurrency rollout collection * Clear error messages instead of crashes when a config file contains invalid YAML ### Documentation * New Build Verifiers section with verification patterns and multi-reward verification * New Evaluate section covering benchmarks, evaluation metrics, and a guide to agent-native results diagnostics * New page for configuring and evaluating agent skills ### Release Assets [GitHub Release v0.4.0](https://github.com/NVIDIA-NeMo/Gym/releases/tag/v0.4.0) #### v0.3.0 ### Release Summary NeMo Gym v0.3.0 ships alongside the NVIDIA Nemotron 3 Ultra model release, open sourcing the environments and corresponding datasets used during training. Highlights: * 70+ new environments, including benchmarks such as Tau2 and Nemotron RL training environments * Popular harness available out-of-the-box such as Claude Code and Hermes * Integrations with OpenEnv and Harbor - use environments from these libraries directly with NeMo Gym * Integration with VeRL - train with VeRL and scale rollout collection with NeMo Gym ### First-Time Contributors We welcomed 30+ new contributors to this release! Here are a few highlights: * **@grace-lam** added the integration to run Harbor environments with NeMo Gym * **@aleksficek** — added Competitive Coding Challenges environment * **@jthomson04** improved rollout resilience when models emit malformed tool-call arguments or missing message content Thank you to all the new contributors for helping make NeMo Gym better! ### New Environments & Benchmarks Added 70+ new environments including novel datasets and integrations of popular benchmarks. New coverage spans: * **Coding** — competitive programming, code infilling, SQL generation, and software-engineering benchmarks with execution-based verification * **Math & proofs** — olympiad-style problems, proof grading and validation, and formal verification (including Lean) * **Knowledge & science** — graduate-level QA, chemistry and physics tasks, and lab-style reasoning (including multimodal figure, table, and protocol tasks) * **Agentic** — multi-turn tool use, search, sandboxed execution, finance workflows, and tau-bench-style conversational agents * **Instruction following** — format constraints, citation compliance, and IFBench-style rule verification * **Safety & RLHF** — jailbreak detection, abstention calibration, prompt-injection resistance, and generative reward modeling * **Multimodal, speech & translation** — VLM benchmarks, visual grounding, ASR evaluation, and machine-translation quality metrics * **Chat & broad knowledge** — arena-style preference evaluation and MMLU-family benchmarks * **Interactive RL** — Gymnasium-style multi-step environments for spatial and game-based training See the [Available Environments](https://github.com/NVIDIA-NeMo/Gym#-available-environments) table for the full list. ### Configure Agent Harnesses * Claude Code — available out of the box in NeMo Gym * Hermes — available out of the box in NeMo Gym * LangGraph agent — an adapter that lets you build custom agents using LangGraph patterns (reflection, subagent orchestration, parallel thinking, rewoo) * Gymnasium agent — generic multi-turn harness for use with OpenAI Gym-style environments ### Configure Models * Optional `max_concurrent_requests` on the OpenAI model server to cap in-flight API calls — useful for rate-limited external endpoints when rollout concurrency is high ### Rollout Collection & Profiling * New `ng_aggregate_rollouts` command to merge rollout shards collected independently across multiple nodes, enabling distributed eval without requiring a single coordinated collection job ### Environment Library Integrations * OpenEnv — combine OpenEnv environments with NeMo Gym environments * Harbor — combine Harbor environments with NeMo Gym environments ### Deprecation Notices * Documentation has moved from Sphinx to Fern. Old Sphinx URLs redirect to the new site at docs.nvidia.com/nemo/gym. The retired `docs/` directory has been removed. ### Bug Fixes * Fixed aiohttp connection limit exhaustion under FastAPI/Uvicorn with multiple workers * Fixed session cookie propagation for Starlette >= 1.0.0 * Fixed duplicated usage counting and errors on empty usage in subsequent model calls * Improved rollout resilience when models emit malformed tool-call arguments or missing message content * Fixed prompt-key hashing when inputs contain Pydantic BaseModel objects ### Documentation * New concepts pages for environments, evaluation, and training * Improved Architecture page to clarify how environments map to NeMo Gym components * Consolidated detailed setup and quickstart into a single improved quickstart with clearer descriptions * Expanded Ecosystem page with environment library, training framework, and agent harness integrations ### Release Assets [GitHub Release v0.3.0](https://github.com/NVIDIA-NeMo/Gym/releases/tag/v0.3.0) #### v0.2.1 Fixed PyPI package distribution that was broken in v0.2.0. No functional changes — all features and fixes from v0.2.0 apply. #### v0.2.0 NeMo Gym v0.2.0 ships alongside the NVIDIA Nemotron 3 Super model release, open sourcing the RL environments and corresponding datasets used during training. This release adds 17 new training environments across coding, math, science, reasoning, agentic tasks, and safety, plus integrations with Aviary, Reasoning Gym, and Verifiers to combine additional environments. You can now run end-to-end rollout collection locally with vLLM and install directly from PyPI. ### New Environments Added 17 new resources servers spanning: * **Coding**: Text to SQL, SWE RL Gen, SWE RL LLM Judge * **Math**: Lean4 Mathematical Proofs * **Science**: Aviary, NewtonBench * **Reasoning**: MultiChallenge, ARC-AGI * **Agent tasks**: xLAM Function Calling, Tavily Search, Single Step Tool Use, Terminus Judge, NeMo Skills Tools * **Safety**: Jailbreak Detection, Over Refusal Detection * **RLHF**: Generative Reward Model Compare Added 5 new agent servers: Aviary agent, proof refinement agent, SWE agents, tool simulation agent, and verifiers agent. **Environment library integrations**: Future House Aviary, Open-Thought Reasoning Gym, Prime Intellect Verifiers. ### Model Serving * Local vLLM model server with end-to-end rollout collection without an external API * vLLM 0.16+ support for the reasoning field in responses * Per-task chat templates and extra body args to support different model configurations across environments in multi-environment training ### Rollout Collection & Profiling * New `ng_reward_profile` command to compute per-task pass rates and aggregate metrics * CPU profiling for rollout performance analysis * Seeding on num\_repeats for reproducible rollouts ### Infrastructure & Developer Experience * PyPI compatibility: install via `pip install nemo-gym` * Dry run mode: `ng_run +dry_run=true` to validate configs and install environments without starting servers * `ng_status` command to list running servers and their health * FastAPI worker support for higher throughput across multiple workers * Server stdout/stderr redirection with server name prefixes ### Model Recipes * Nemotron 3 Nano 30B end-to-end training recipe with single-GPU and multi-node tutorials ### Documentation * Added training tutorials for Unsloth, TRL, and Nemotron 3 Nano (single-GPU and multi-node) * Added environment tutorials for creating environments, custom data preparation, and integrating external libraries * Rewrote concepts documentation with new training approaches page, architecture diagrams, and expanded agent/resources server docs * Revamped ecosystem page with training framework and environment library integrations * Added deployment topology and SWE RL infrastructure case study * Site-wide quality sweep: consistent naming, style guide, redirects, and FAQ additions ### Bug Fixes * Fixed 0.1.1 environments to work correctly with RL training pipelines * Fixed crash when server receives malformed JSON during rollout collection * Fixed dry run mode failing after initial implementation * Fixed nested `responses_create_params` overrides not merging correctly from CLI * Fixed `ng_prepare_data` failing when multiple environments define overlapping metrics * Fixed reward profiling failing when model response doesn't include usage stats * Fixed NeMo-Skills python tool to use HTTP calls instead of subprocess execution * Bumped Pillow and other packages to address security vulnerabilities * `ng_dump_config` now redacts API key values from output ### First-Time Contributors We'd like to highlight the following first-time contributors: * **@sidnarayanan** added the Aviary integration to enable training on any Aviary environment, a library of interactive RL environments spanning math, science, biology, and more * **@3mei** added the text-to-SQL environment to generate SQL queries from natural language across multiple SQL dialects * **@Kelvin0110** added the NewtonBench environment to discover scientific laws through interactive experimentation #### v0.1.1 Initial public release of NeMo Gym.