datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
R2E-Gym-Subset-Verified
R2E-Gym-Subset-Verified
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 the gold patch, run the
upstream /testbed/run_tests.sh baked into the row's image, check the parsed outcomes against
expected_output_json.
Changes vs upstream
Validation-only subset — our passes, run in fresh sandboxes per row: one full pass at
concurrency 200, then a 10× retry pass over… See the full description on the dataset page: https://huggingface.co/datasets/PrimeIntellect/R2E-Gym-Subset-Verified.CUA-Gym
CUA-Gym
CUA-Gym is a collection of verifiable computer-use agent tasks for reinforcement learning with verifiable rewards (RLVR). Each task pairs a natural-language instruction with executable setup artifacts and a Python reward function that checks task completion programmatically. For details, see the paper CUA-Gym: Scaling Verifiable Training Environments and Tasks for Computer-Use Agents.
This release contains the full public CUA-Gym task set after the necessary data review.… See the full description on the dataset page: https://huggingface.co/datasets/xlangai/CUA-Gym.R2E-Gym-Full
R2E-Gym Subset Filtered for MAGRPO
Filtered subset of R2E-Gym optimized for 2-agent MAGRPO training with 7B models.
Dataset Statistics
Total instances: 167
Format: Issue description + Oracle files in prompt
Optimized for: 2-agent collaboration, 7B models
Filtering Criteria (SWE-bench Lite Style)
Problem statement: >40 words (up to 500 for context window)
Must have non-empty oracle patch (non-test file changes)
File count: Exactly 1 oracle file (single-file… See the full description on the dataset page: https://huggingface.co/datasets/ryankamiri/R2E-Gym-Full.Muse-Glimmer-SWE-Gym-2k
Muse-Glimmer-SWE-Gym-2k
Agentic coding traces from meta-models/Muse-Glimmer-30B, recorded for training a
speculative-decoding drafter. 1,981 mini-swe-agent trajectories over SWE-Gym and
SWE-bench-extra instances, and the 159,999 individual chat-completion calls behind them.
Configs
Config
Rows
Size
What it is
train
1,981
57 MB
One row per trajectory: the full conversation as messages.
raw
159,999
2.7 GB
One row per recorded API call: request and… See the full description on the dataset page: https://huggingface.co/datasets/Satgoy152/Muse-Glimmer-SWE-Gym-2k.Gym_Salesman_Dataset
Gym Salesman Dataset
11,997 synthetic gym-membership sales conversations, each labelled SUCCESS or FAILURE.
🔗 Project Links
| Live App — practice against an AI customer | Hugging Face Space |
| Telegram Bot — practice on the go | @ido_salescoach_bot |
| Dataset — 11,997 labelled conversations | elg4/Gym_Salesman_Dataset |
| Data Generation — how the data was built | notebook |
| Recommendation — the embedding retriever | notebook |
Every conversation is a… See the full description on the dataset page: https://huggingface.co/datasets/elg4/Gym_Salesman_Dataset.multilingual-reasoning-gym-sft
Reasoning Gym SFT Dataset
This dataset contains Supervised Fine-Tuning (SFT) reasoning data procedurally generated using Reasoning Gym environments.
It is designed to train reasoning models (such as DeepSeek-R1-style or Qwen-Coder-style models) to explain their step-by-step reasoning chain before outputting a final answer wrapped inside LaTeX \boxed{...}.
Where Does This Dataset Come From?
This dataset is procedurally generated from Reasoning Gym, an open-source… See the full description on the dataset page: https://huggingface.co/datasets/MauroPello/multilingual-reasoning-gym-sft.R2E-Gym
R2E-Gym: OpenCode, Codex, and Claude Code environments
This repository contains three environment variants of
R2E-Gym/R2E-Gym-Subset.
Each variant contains the same 4,578 tasks in five Parquet shards, with the
original 14-column schema and task order. Only docker_image is replaced with
the corresponding public image containing the selected agent.
Configuration
Files
Docker image prefix
opencode
opencode/data/*.parquet
docker.io/loongsage/r2e-gym:opencode_
codex… See the full description on the dataset page: https://huggingface.co/datasets/loongsage/R2E-Gym.R2E-Gym-Subset
R2E-Gym Subset Filtered for MAGRPO
Filtered subset of R2E-Gym optimized for 2-agent MAGRPO training with 7B models.
Dataset Statistics
Total instances: 54
Format: Issue description + Oracle files in prompt
Optimized for: 2-agent collaboration, 7B models
Filtering Criteria (SWE-bench Lite Style)
Problem statement: >40 words (up to 500 for context window)
Must have non-empty oracle patch (non-test file changes)
File count: Exactly 1 oracle file (single-file… See the full description on the dataset page: https://huggingface.co/datasets/ryankamiri/R2E-Gym-Subset.ppl-gym-rollouts
ppl-gym rollouts
Model rollouts from ppl-gym, a benchmark of probabilistic-programming
problems: language-neutral problem statements with per-language ground-truth
realizations. Each rollout is one LLM-generated solution (program) for a
(model, language, problem) triple, executed and scored against ground truth
under a single answer comparator.
Code & methodology: https://github.com/shepardxia/ppl-gym
Browser: https://pplgym.kingdomofends.org
Configs
rollouts… See the full description on the dataset page: https://huggingface.co/datasets/Sheppp/ppl-gym-rollouts.Nemotron-RL-IF-Gym-LFM2.5-prompts
Nemotron-RL instruction-following & abstention Gym tasks → LFM2.5 prompt format
Verifiers are NVIDIA NeMo-Gym rule checkers, not IFEval. Each row keeps the full verifier metadata in extra (verifier regex/string-match specs, schema_str/schema_type, exp_cal_state, answer) so the checks can be re-implemented or run through NeMo Gym's resources servers. qa_abstention rows carry a per-row license of CC BY-SA 4.0 from the source.
A derivative of [five nvidia/Nemotron-RL-* NeMo-Gym… See the full description on the dataset page: https://huggingface.co/datasets/hungbenjamin402/Nemotron-RL-IF-Gym-LFM2.5-prompts.hybrid_gym_func_gen_raw
Hybrid Gym: Function Generation Dataset
Dataset for the Function Generation benchmark task. Each instance contains a function with its signature and docstring, where the agent must implement the function body.
See the benchmark README for usage instructions.
R2E-Gym-Collab-v3
R2E-Gym-Collab-v3
A high-quality subset of R2E-Gym optimized for multi-agent RL training (MAGRPO).
What's Special
This dataset is filtered for optimal gradient signal in Level 3 (test execution) rewards:
2-20 failing tests per instance (good gradient signal)
Collaboration-suitable (AI-verified two-agent task decomposition)
Test failures in prompt (explicit error context for the model)
Why This Filtering?
Test Count
L3 Reward per Fix
Problem
1… See the full description on the dataset page: https://huggingface.co/datasets/ryankamiri/R2E-Gym-Collab-v3.R2E-Gym-Collab-v2
R2E-Gym-Collab-v2
This dataset extends ryankamiri/R2E-Gym-Collab by adding test failure information to each prompt.
What's New
Each prompt now includes a [TEST_FAILURES] section showing:
The actual pytest output from failing tests (before the fix)
Error messages, assertion failures, and tracebacks
This gives models context about WHAT needs to be fixed, not just WHERE
Prompt Structure
[ISSUE]
Bug description, expected vs actual behavior
[/ISSUE]… See the full description on the dataset page: https://huggingface.co/datasets/ryankamiri/R2E-Gym-Collab-v2.hybrid_gym_func_localize_raw
Hybrid Gym: Function Localization Dataset
Dataset for the Function Localization benchmark task. Each instance contains a function description (without file path or name), and the agent must locate the function in the repository and add a docstring.
See the benchmark README for usage instructions.
Fields
instance_id: Unique identifier
repo: GitHub repository (owner/repo)
base_commit: Commit hash to checkout
file_path: Path to the file containing the target function… See the full description on the dataset page: https://huggingface.co/datasets/hybrid-gym/hybrid_gym_func_localize_raw.hybrid_gym_dep_search_raw
Hybrid Gym: Dependency Search Dataset
Dataset for the Dependency Search benchmark task. Each instance contains a target function and its ground-truth dependencies (functions/classes directly called by that function) within a repository.
See the benchmark README for usage instructions.
Fields
instance_id: Unique identifier
repo: GitHub repository (owner/repo)
base_commit: Commit hash to checkout
target_function_name: Name of the target function
target_function_file: Path… See the full description on the dataset page: https://huggingface.co/datasets/hybrid-gym/hybrid_gym_dep_search_raw.