RL environments on the Hugging Face Hub
Explore reinforcement learning environments and tasks on the Hugging Face Hub. Browse OpenEnv, Harbor, MiMo, NeMo Gym and Verifiers, inspect rewards, and run supported agent rollouts.
- Agentic RL Environments (XiaomiMiMo/MiMo-V2.6-RL-oss, MiMo RL release)
- BlenderWorld (Qwen/BlenderWorld, Harbor dataset) BlenderWorld is the first comprehensive benchmark dedicated to evaluating frontier models on realistic 3D creation tasks. BlenderWorld comprises 223 tasksā¦
- openai/math as Harbor environments (FineEnvs/openai-math, Harbor dataset) Experimental. Built on a research release that is itself new and not yet independently reviewed. Expect rough edges, and read any proof that scores 1 beforeā¦
- Terminal-Bench 3.0 (harborframework/terminal-bench-3.0, Harbor dataset)
- Terminal-Bench (harborframework/terminal-bench, Harbor dataset)
- š SmolDataEnvs (FineEnvs/SmolDataEnvs, RL dataset) 5.5K+ RL tasks for hill-climbing small models in code and data science.
- Terminal-Bench 2.1 (Harbor git-repos dataset) (harborframework/terminal-bench-2.1, Harbor dataset)
- Terminal-Bench-Science (harborframework/terminal-bench-science, Harbor dataset)
- ExtractBench (llamaindex/ExtractBench, Harbor dataset) Quick links: š Website š Paper š» Code Given a document and a schema, a system returns structured data with evidence. The input is a full documentā¦
- Zhanli-Li/DataWorld-Train-2k-part1 (Zhanli-Li/DataWorld-Train-2k-part1, Harbor dataset) Scaling Data Environments from Scratch Specify the world behind the files. Generate the data. Ground the supervision.
- TerminalHorizon-Environment (shunzou05/TerminalHorizon-Environment, Harbor dataset) Scaling Agentic Data for Long-Horizon Terminal Intelligence Paper Ā· Project Page Ā· Citation TerminalHorizon is a fully automated data synthesis engine forā¦
- ExtractBench Pro (llamaindex/ExtractBenchPro, Harbor dataset) Quick links: š Website š» Code š ExtractBench ExtractBench Pro is the hard companion to ExtractBench. The task is the same: given a full document and aā¦
- nvidia/Nemotron-RL-Agentic-Conversational-Tool-Use-Pivot-v1 (nvidia/Nemotron-RL-Agentic-Conversational-Tool-Use-Pivot-v1, NeMo Gym dataset) We created an RL dataset for conversational tool-use by utilizing existing expert tool-use trajectories. We pose each assistant step of the trajectory as aā¦
- nvidia/Nemotron-RL-coding-competitive_coding (nvidia/Nemotron-RL-coding-competitive_coding, NeMo Gym dataset) The Nemotron-RL-coding-competitive coding dataset is a python-only, reasoning-based, synthetic dataset. It contains competitive coding style problems andā¦
- SkillTransferBench (NJU-LINK/SkillTransferBench, RL dataset) 35 task families Ā· 408 executable cases Ā· 6 domains
- harborframework/terminal-bench-2.0 (harborframework/terminal-bench-2.0, Harbor dataset)
- AgentTrove (open-thoughts/AgentTrove, Harbor dataset) AgentTrove is the largest open-source collection of agentic interaction traces to date, released by the OpenThoughts-Agent team. It contains 1,696,847 rowsā¦
- Unslop (Harbor) (rawmodels/unslop-harbor, Harbor dataset) 200 Harbor tasks with the reward we trained AlexWortega/Qwen3.8-27B-antislop on. Your model finishes a tweet or writes a HumanEval function body. Claude Opusā¦
- APEX-Accounting ā Harbor edition (10-task dev set) (mercor/apex-accounting-harbor, Harbor dataset) APEX-Accounting is a benchmark built by Mercor in partnership with Ramp to assess whether frontier models can do the real work of accountants: reconcilingā¦
- lingee-bench (KD-Lingee/lingee-bench, Harbor dataset) An enterprise-grade agent benchmark developed by Kingdee. Tasks use the native Harbor format: each task is a directory containing the instruction, sandboxā¦
- MiMo-V2.6-RL General (Harbor) (FineEnvs/MiMo-V2.6-RL-harbor-general, Harbor dataset) Work in a simulated company through its MCP systems. 925 Harbor tasks from the General domain of Xiaomi's MiMo-V2.6-RL-oss, the RL environments MiMo-V2.6 wasā¦
- Lego-X/LegoFlow-SWE (Lego-X/LegoFlow-SWE, Harbor dataset) LegoFlow-SWE Ā· 5,000 verified Harbor SWE tasks and two GLM-5.2 trajectory releases
- nvidia/Nemotron-RL-agent-workplace_assistant (nvidia/Nemotron-RL-agent-workplace_assistant, NeMo Gym dataset) The Nemotron-RL-agent-workplace assistant is a tool use - multi step agentic environment that tests the agentās ability to execute tasks in a workplaceā¦
- Procedural Pile (reasoning-core/procedural-pile, Verifiers environment) Verifiable reasoning problems, generated by code, formatted as supervised training data. Procedural Pile contains 24,806,238 problems from 55 task familiesā¦
- Zhanli-Li/DataWorld-Train-2k-part2 (Zhanli-Li/DataWorld-Train-2k-part2, Harbor dataset) Scaling Data Environments from Scratch Specify the world behind the files. Generate the data. Ground the supervision.
- Calendar-Scheduling-Dataset (nvidia/Nemotron-RL-agent-calendar_scheduling, NeMo Gym dataset) The Calendar-Scheduling-Dataset is a multi-turn conversation dataset that can understand natural language scheduling constraints, follow instructions acrossā¦
- Unified Code VERL Dataset (sungyub/code-verl-unified, RL dataset) This dataset aggregates seven code-reasoning collections into a single VERL-formatted repository containing approximately 958,539 unique problems. Theā¦
- MiMo-V2.6-RL Code (Harbor) (openmodelai/MiMo-V2.6-Code, Harbor dataset) Fix a real issue in a real repository. 2,698 Harbor tasks from the Code domain of Xiaomi's MiMo-V2.6-RL-oss, the RL environments MiMo-V2.6 was trained onā¦
- Agent World Model in the Wild (ChilleD/WildAWM, Harbor dataset) Agent World Model in the Wild (WildAWM) releases 341 accepted cross-interface (cli/mcp/web) tasks over frozen persona-day worlds in Harbor task format, splitā¦
- directus (earakely-scale/DirectusEnv-REA, RL dataset) directus An AgentEnv bundle with its recorded runs, published with agentenv-hf.
- espocrm (earakely-scale/EspoCRMEnv-REA, RL dataset) espocrm An AgentEnv bundle with its recorded runs, published with agentenv-hf.
- shlink (earakely-scale/ShlinkEnv-REA, RL dataset) shlink An AgentEnv bundle with its recorded runs, published with agentenv-hf.
- SplatBench (perpetual-research/splatbench, RL dataset) SplatBench is a benchmark and RL environment for raw spatial reasoning. A language model gets eight photos of an object and rebuilds the object in 3D byā¦
- SWE-Lego-Real-Data (PrimeIntellect/SWE-Lego-Real-Data, Verifiers environment) Fixed fork of SWE-Lego/SWE-Lego-Real-Data (paper): 4,432 / 5,009 resolved real-GitHub-issue tasks (Python) that can actually be scored. For the additionallyā¦
- Math-VERL Unified Collection (sungyub/math-verl-unified, RL dataset) A unified collection of 9 high-quality mathematical reasoning datasets totaling 2,269,166 deduplicated problems, all converted to VERL format forā¦
- Agent Workbench (stoiclion/agent-workbench, RL dataset) An OpenEnv environment of short, tool-using work tasks across eight domains: finance, industrial, math, media, office, science, security, software. Imageā¦
- Quant Desk ā OpenEnv Arena dataset (Akahsizrr/quant-desk, RL dataset) Data for the Quant Desk OpenEnv environment: 36 deterministic, self-contained single-answer tasks spanning all eight Arena domains (software engineeringā¦
- Unspoken hackathon data (AumPandya01/unspoken-data, NeMo Gym dataset)
- Concordance Office and Industrial (Codcordance/concordance-office-industrial-v1, Harbor dataset) Eight original tasks reused from vibrantlabsai/Pool Desk v5, pinned at c5ea7ae8510fe2f01ee9c3154f6f6af0317bdead. This dataset preserves the selectedā¦
- rStar-Coder-VERL Dataset (sungyub/rstar-coder-verl, RL dataset) This dataset contains 386,640 coding problems from the microsoft/rStar-Coder collection (synthetic rl testcase configuration), transformed to the VERLā¦
- Concordance Records (Codcordance/concordance-records-v1, RL dataset) Twelve terminal tasks for reconciling records, across six original families. Each family has two versions. All records are simulated; no peer-authoredā¦
- Mixed Skills v13: eight-domain verified workspace tasks (cmunley/mixed-skills-env-v13, RL dataset) One OpenEnv server with deterministic procedurally generated tasks across the eight evaluation domains: software engineering, industrial and physical systemsā¦
- Mixed Skills v12: eight-domain verified workspace tasks (cmunley/mixed-skills-env-v12, RL dataset) One OpenEnv server, 406 deterministic procedurally generated tasks across the eight evaluation domains: software engineering, industrial and physical systemsā¦
- Skills Desk (Syndicate604/skills-desk-v1, RL dataset) Frozen v1 copy. Identical to Syndicate604/skills-desk at revision a785ccdc9784ed863d0ffecec826d72176c0a785 (40 training tasks, imageā¦
- Geoguesser Environment (FineEnvs/geoguesser-env, OpenEnv Space)
- PortSimEnv v1 (FineEnvs/PortSimEnv, OpenEnv Space) Dock planning on real Port of Barcelona calls
- TabulaClean (lilpookie404/tabulaclean, OpenEnv Space)
- PrimeIntellect/Terminal-Lego-15k (PrimeIntellect/Terminal-Lego-15k, Harbor dataset)
- Spreadsheet-RL Dataset (Spreadsheet-RL/Spreadsheet-RL, RL dataset) This dataset contains the training and evaluation data used by Spreadsheet-RL, a reinforcement learning framework for spreadsheet agents that edit Excelā¦
- Long-Horizon Terminal-Bench (LHTB) (IntelligenceLab/Long-Horizon-Terminal-Bench, Harbor dataset) Leaderboard: the panel at the top of this page has two views. Use the Task menu in its top-right corner to switch between Lhtb2 Solved (LHTB 2.0, out of 77ā¦
- MiMo-V2.6-RL Code (Harbor) (FineEnvs/MiMo-V2.6-RL-harbor-code, Harbor dataset) Fix a real issue in a real repository. 2,698 Harbor tasks from the Code domain of Xiaomi's MiMo-V2.6-RL-oss, the RL environments MiMo-V2.6 was trained onā¦
- š SmolDataEnvs: Harbor (train) (FineEnvs/SmolDataEnvs-harbor-train, Harbor dataset) 5.5K+ RL tasks for hill-climbing small models in code and data science.
- Multi-SWE-bench (PrimeIntellect/Multi-SWE-bench, Verifiers environment) Re-upload of ByteDance's Multi-SWE-bench evaluation benchmark: 2,132 issue-resolving tasks across the seven Multi-SWE languages. This is the held-out evalā¦
- š§© ProgramGym-1648 (ProgramGym/programgym-1648, Harbor dataset) A dataset of 1,648 tasks for long-horizon program reproduction agents.
- Reverse-Text-RL (PrimeIntellect/Reverse-Text-RL, Verifiers environment) A small, scrappy RL dataset used in prime-rl's CI to debug RL training asking a model to reverse small sentences character-by-character. Follows the generalā¦
- TaskTrove (open-thoughts/TaskTrove, Harbor dataset) v5.1 (current) ā independent-review source retirement ā moves 15 sources with majority or unanimous REJECT verdicts out of the default config and intoā¦
- MiMo-V2.6-RL Terminal (Harbor) (FineEnvs/MiMo-V2.6-RL-harbor-terminal, Harbor dataset) Terminal-Bench style tasks. 64 Harbor tasks from the Terminal domain of Xiaomi's MiMo-V2.6-RL-oss, the RL environments MiMo-V2.6 was trained on, converted soā¦
- Monthly-SWEBench 2026-05 (UnipatAI/Monthly-SWEBench-2026-05, Harbor dataset) This package contains the 2026-05 Monthly-SWEBench final release set. It includes 100 Harbor-format software engineering tasks selected from closed GitHub PRsā¦
- Monthly-SWEBench-2026-04 (UnipatAI/Monthly-SWEBench-2026-04, Harbor dataset) 90 tasks ā 43 bugfix + 47 non-bugfix Tasks span diverse open-source repositories Each task includes a runnable environment, test suite, and reference solution
- Monthly-SWEBench-2026-03 (UnipatAI/Monthly-SWEBench-2026-03, Harbor dataset) 112 tasks ā 68 bugfix + 44 non-bugfix Tasks span diverse open-source repositories Each task includes a runnable environment, test suite, and reference solution
- PortSimEnv on AgentEnv (earakely-scale/PortSimEnv-AgentEnv, RL dataset) GPT-6 Astra plays the lines week dock-24B-w07x1-busy-0-e09, APM Terminals Barcelona's busy week of 12 February 2024 in the wind of the storm week of 11ā¦
- Human Behavior Atlas (HumanBehaviorAtlas/human_behavior_atlas, RL dataset) A large-scale multimodal dataset for human behavior understanding, spanning emotion recognition, sentiment analysis, humor detection, mental health screeningā¦
- MiMo-V2.6-RL Cyber (Harbor) (FineEnvs/MiMo-V2.6-RL-harbor-cyber, Harbor dataset) Reproduce a real memory-safety crash (ARVO). 1,000 Harbor tasks from the Cyber domain of Xiaomi's MiMo-V2.6-RL-oss, the RL environments MiMo-V2.6 was trainedā¦
- TaskTrove Clean (open-athena/task-trove, Harbor dataset) All 12,720 tasks in laion swesmith-oracle-filtered-v2 now install the verifyit grading package at Marin commit a9adee624d703250556f8078b3d5beb18b4b57ab. Thisā¦
- š Data Agent ā Harbor (train) (FineEnvs/data-agent-harbor-train, Harbor dataset) Teach an agent to actually do data science. This is a suite of 5,000 hands-on data-analysis tasks: each one drops your agent into a sandbox with a realā¦
- HF ML Tasksmith (FineEnvs/HF_ML_Tasksmith, Harbor dataset) Fifty PR-derived Harbor tasks from Accelerate, Diffusers, PEFT, Transformers and TRL, including CPU and GPU tasks. Contains 50 Harbor tasks generated with theā¦
- MiMo-V2.6-RL Webdev (Harbor) (FineEnvs/MiMo-V2.6-RL-harbor-webdev, Harbor dataset) Build a website from a design brief. 2,093 Harbor tasks from the Webdev domain of Xiaomi's MiMo-V2.6-RL-oss, the RL environments MiMo-V2.6 was trained onā¦
- MiMo-V2.6-RL Music (Harbor) (FineEnvs/MiMo-V2.6-RL-harbor-music, Harbor dataset) Compose a piece in ABC notation. 1,000 Harbor tasks from the Music domain of Xiaomi's MiMo-V2.6-RL-oss, the RL environments MiMo-V2.6 was trained onā¦
- Monkey Business (ScalingIntelligence/monkey_business, Verifiers environment) Monkey Business is a dataset of samples from large language models. It contains both correct and incorrect samples from a variety of models (the Llama-3ā¦
- WildClawBench-Harbor (internlm/WildClawBench-Harbor, Harbor dataset) This repository is the WildClawBench benchmark converted to the Harbor task format, so that all 60 tasks can be run directly with harbor run against anyā¦
- LHTB Leaderboard ā Long-Horizon Terminal-Bench (IntelligenceLab/LHTB-leaderboard, Harbor dataset) This repository hosts submitted runs for Long-Horizon Terminal-Bench (LHTB), a 46-task benchmark measuring how well LLM agents sustain useful work in aā¦
- ProgramDistill-Bench (microsoft/ProgramDistill, Harbor dataset) ProgramDistill is a dataset and executable benchmark for evaluating whether coding agents can restore missing behavior in interactive web applications. Eachā¦
- Repo2RLEnv TerminalWorld (EuniAI) (FineEnvs/repo2rlenv-terminalworld, Harbor dataset) Screen original public terminal recordings, score feasibility and value, extract and refine a solution, write its instruction, reconstruct an offlineā¦
- R2E-Gym-Subset-Verified (PrimeIntellect/R2E-Gym-Subset-Verified, Verifiers environment) Gold-patch-validated subset of R2E-Gym/R2E-Gym-Subset (paper). The train split contains 4,522 / 4,578 rows (98.78%) verified scoreable end-to-end: apply theā¦
- š SmolDataEnvs: Harbor (test) (FineEnvs/SmolDataEnvs-harbor-test, Harbor dataset) 5.5K+ RL tasks for hill-climbing small models in code and data science.
- MegaTerminal (di-zhang-fdu/MegaTerminal, Harbor dataset) MegaTerminal is a Harbor-style terminal task-folder dataset assembled from the terminal task sources collected for task matching experiments. The task foldersā¦
- Repo2RLEnv R2E (FineEnvs/repo2rlenv-r2e, Harbor dataset) Author equivalence tests around repository functions, measure their execution and coverage remotely, and export function-reconstruction tasks with theā¦
- dabstep-harbor (AdithyaSK/dabstep-harbor, Harbor dataset) A /workdir-native Harbor edition of DABstep. 450 financial-analytics factoid tasks over a single shared Adyen-payments corpus ā every task's workingā¦
- repo2rlenv-pr-runtime (FineEnvs/repo2rlenv-pr-runtime, Harbor dataset) Generated by Repo2RLEnv ā turning real GitHub repositories into verifiable RL environments.
- Repo2RLEnv SWE-gen (FineEnvs/repo2rlenv-swe-gen, Harbor dataset)
- Repo2RLEnv Endless Terminals (FineEnvs/repo2rlenv-endless-terminals, Harbor dataset) Sample native terminal categories, complexity and scenarios; author instructions and private expected behavior, initial-state tests, final tests, fixtures andā¦
- Repo2RLEnv SWE-Next (FineEnvs/repo2rlenv-swe-next, Harbor dataset) Mine real repository history for supported code changes and associated tests, preserve regression behavior, and export the historical implementation change asā¦
- repo2rlenv-pr-diff (FineEnvs/repo2rlenv-pr-diff, Harbor dataset) Generated by Repo2RLEnv ā turning real GitHub repositories into verifiable RL environments.
- repo2rlenv-commit-runtime-v2 (FineEnvs/repo2rlenv-commit-runtime, Harbor dataset) Generated by Repo2RLEnv ā turning real GitHub repositories into verifiable RL environments.
- š§Ŗ Data Agent ā Harbor (eval) (FineEnvs/data-agent-harbor-eval, Harbor dataset) A small, difficulty-balanced validation split ā 144 tasks ā perfect for quick checkpoints while you train. Same idea as the rest of the family: your agentā¦
- Multi-SWE-RL-Verified (PrimeIntellect/Multi-SWE-RL-Verified, Verifiers environment) Gold-patch-validated subset of PrimeIntellect/Multi-SWE-RL-Reupload (ByteDance's Multi-SWE-RL): 2,232 / 4,703 rows across C, Go, Java, JavaScript, Rust, andā¦
- š SmolDataEnvs: Harbor (eval) (FineEnvs/SmolDataEnvs-harbor-eval, Harbor dataset) 5.5K+ RL tasks for hill-climbing small models in code and data science.
- šÆ Data Agent ā Harbor (test) (FineEnvs/data-agent-harbor-test, Harbor dataset) A held-out benchmark for data-analysis agents: 250 tasks, deliberately balanced across difficulty and leaning toward the harder end so it actually separatesā¦
- Repo2RLEnv SWE-smith (FineEnvs/repo2rlenv-swe-smith, Harbor dataset) Apply owned procedural mutations to real repository functions, require an executable test contrast, and author a natural issue from bounded evidence. Containsā¦
- Scale-SWE-Verified (PrimeIntellect/Scale-SWE-Verified, Verifiers environment) Gold-patch-validated fork of AweAI-Team/Scale-SWE (paper): 17,202 / 20,181 Python issue-resolving tasks that produce a clean reward signal end-to-end. Defaultā¦
- Repo2RLEnv SCALER (FineEnvs/repo2rlenv-scaler, Harbor dataset) Generate instances from owned reasoning families with seeded procedural samplers, deterministic references and family-specific answer verification; generationā¦
- Repo2RLEnv TMax (FineEnvs/repo2rlenv-tmax, Harbor dataset) Sample the native TMax legacy taxonomy, create a template, initial-state tests and final completion tests in distinct stages, then build fixtures and aā¦
- Repo2RLEnv SETA Evol (FineEnvs/repo2rlenv-seta-evol, Harbor dataset) Start with complete Harbor parents, apply one of six named evolution strategies, then author a complete child task with its own instruction, environmentā¦
- Repo2RLEnv SETA Seed2Synth (FineEnvs/repo2rlenv-seta-seed2synth, Harbor dataset) Use attributed Unix Stack Exchange questions and accepted answers to design terminal tasks, then author self-contained offline environments, instructionsā¦
- Repo2RLEnv DataArc terminal synthesis (Envs-FORGE-linked code) (FineEnvs/repo2rlenv-dataarc, Harbor dataset) Read complete Harbor parent tasks and enumerate few-shot, self-instruct and evolution transformations. Author a complete child environment, instructionā¦
- PatchRecoveryGym for Laguna (poolside-laguna-hackathon/patchrecoverygym-laguna, Verifiers environment) Submitted by: Kannappan Sirchabesan (@kannappans) Ā· Poolside Research Hackathon (Foundations track) A reproducible eval + RL environment that tests whether aā¦
- Repo2RLEnv SWE-Flow (FineEnvs/repo2rlenv-swe-flow, Harbor dataset) Trace passing repository tests, select exercised source functions, remove their implementations and generate a reconstruction instruction. Restore theā¦
- Repo2RLEnv R2E-Gym / SWEGEN (FineEnvs/repo2rlenv-r2e-gym, Harbor dataset) Mine supported bug-fix commits, measure failing-to-passing and regression tests across revisions, and export the buggy starting state and deterministicā¦
- earakely-scale/AirOpsEnv-AgentEnv (earakely-scale/AirOpsEnv-AgentEnv, RL dataset) Three GPT-6.1 Sol desks run JetBlue's network through the storm of Tuesday 8 July 2025 (b6-20250708-0400), shown at Kennedy: 637 flights from 04:00 withā¦
- repo2rlenv-code-instruct (FineEnvs/repo2rlenv-code-instruct, Harbor dataset) Generated by Repo2RLEnv ā turning real GitHub repositories into verifiable RL environments.
- repo2rlenv-equivalence-tests (FineEnvs/repo2rlenv-equivalence-tests, Harbor dataset) Generated by Repo2RLEnv ā turning real GitHub repositories into verifiable RL environments.
- Repo2RLEnv CLI-Gym (FineEnvs/repo2rlenv-cli-gym, Harbor dataset) Start with a healthy repository, synthesize a disruption and recovery, verify that damage breaks the original tests and recovery restores them, then export anā¦
- AdithyaSK/data_agent_rl_environment_train_multireward (AdithyaSK/data_agent_rl_environment_train_multireward, Harbor dataset) Multi-reward variant of AdithyaSK/data agent rl environment train (2238 tasks, identical data/instructions). The only change: each task's verifier now emits aā¦
- FACET-Terminal-Tasks-6k (FACET-Terminal/FACET-Terminal-Tasks-6k, Harbor dataset) 6,020 execution-grounded tasks for terminal agents, coding agents, and executable workflow research
- TerminalWorld Seeds, oracle-validated (andylizf/TerminalWorld-Seeds-Clean, Harbor dataset) The current TW selection retains 405 tasks, excluding any original TW task excluded by either the preserved local quality curation or the upstream v3 trainingā¦
- SWE-Lego-Real-Data-Verified (PrimeIntellect/SWE-Lego-Real-Data-Verified, Verifiers environment) Gold-patch-validated subset of PrimeIntellect/SWE-Lego-Real-Data (itself a fixed fork of SWE-Lego's real-data split). The resolved split contains 4,323 /ā¦
- nvidia/Nemotron-RL-knowledge-mcqa (nvidia/Nemotron-RL-knowledge-mcqa, NeMo Gym dataset)
- Terminal-Bench Hard (Zhongzhi1228/Terminal-Bench-Hard, Harbor dataset) Terminal-Bench Hard is a set of 100 challenging terminal-based agent tasks. The tasks cover software engineering, debugging, data processing, systemā¦
- TermGrade Environments: 1,004 executable Linux tasks (ai-and/termgrade-environments, Harbor dataset) Part of TermGrade: graded environments and trajectories for terminal agents. Read the blog post. 1,004 executable environments Ā· graded by 6 models Ā· measuredā¦
- MegaTerminal-Hard-1K (jwu323/MegaTerminal-Hard-1K, Harbor dataset) The 1,000 highest-quality, hardest terminal-agent tasks selected from the 11,599 tasks in di-zhang-fdu/MegaTerminal. Every task is a complete, self-containedā¦
- Watercolour reference pool (FineEnvs/watercolour-reference-pool, RL dataset) The reference paintings that define the reward in the watercolour RL environment: an agent writes a p5.brush sketch, the sketch is rendered, and a visionā¦
- Terminal-Bench 2.1 (offline Apptainer, v1) (laion/terminal-bench-2-1-apptainer-v1, Harbor dataset) A validated subset of Terminal-Bench 2.1 (revision 7131e4375048a0e408a8fb404b5f499d726b695b, Apache-2.0) for running on HPC clusters without Docker andā¦
- repo2rlenv-cve-patches (FineEnvs/repo2rlenv-cve-patches, Harbor dataset) Generated by Repo2RLEnv ā turning real GitHub repositories into verifiable RL environments.
- Warcraft III (wc3env) on AgentEnv (earakely-scale/wc3env-AgentEnv, RL dataset) DeepSeek V4.1 Flash's Undead army meets an identical one played by Warcraft's own attack-move. The plan it wrote at that moment is over the fight. It wins inā¦
- SWE-rebench-V2-Filtered-Verified (PrimeIntellect/SWE-rebench-V2-Filtered-Verified, Verifiers environment) Filtered and gold-patch-verified subset of Nebius's SWE-rebench-V2 (paper): 6,272 / 32,079 freshly-mined GitHub PR tasks across 17 languages. Default datasetā¦
- š Agent Harbor Tasks (efederici/agent-harbor-tasks, Harbor dataset) 400 original shell-agent tasks, Italian and English, from a 4-call chore to a 20-call investigation.
- LoLBench (lolbench26/LoLBench, Harbor dataset) Software development often begins with user intent and high-level design rather than an implementation-ready specification. Developers must ground those goalsā¦
- Nemotron-RL-Agentic-Indirect-Prompt-Injection-v1 (nvidia/Nemotron-RL-Agentic-Indirect-Prompt-Injection-v1, NeMo Gym dataset) Nemotron-RL-Agentic-Indirect-Prompt-Injection-v1 is an RL dataset for training and evaluating a tool-using agent's ability to resist Indirect Prompt Injectionā¦
- SWE-bench Science (OpenMOSS-Team/SWE-bench-Science, Harbor dataset) SWE-bench Science evaluates coding agents on software-engineering tasks drawn from scientific-computing repositories. The release contains 119 tasks across 20ā¦
- SkillFlow Test Tasks (zhang-ziao/SkillFlow-Task, Harbor dataset) SkillFlow Test Tasks is the task repository used in the SkillFlow benchmark for evaluating lifelong skill discovery, skill revision, and cross-task proceduralā¦