datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
rlvr-reward-hacking-scale-no-conftest-20260909-completion
Matched no-conftest RLVR study 20260909-completion
Lossless research records, grouped by model and trajectory type. Only the listed
configurations have published records. Canary diagnostics are excluded from study
estimates; run status in provenance distinguishes retired diagnostics from active
or completed training. Valid failures, refusals and truncations are retained.
The train split name is a dataset-loader convention; record_type identifies
whether a record is training… See the full description on the dataset page: https://huggingface.co/datasets/lucabaroni/rlvr-reward-hacking-scale-no-conftest-20260909-completion.Lego-RL-2699
SWE-Lego-RL-2699
2,699 executable, difficulty-filtered SWE tasks for agentic RL, shipped in two
parallel views of the same instances:
View
Path
What it is
Official OpenSWE records
openswe_official_2699/
The original upstream GAIR/OpenSWE rows for exactly these 2,699 instances
Harbor RL environments
openswe_harbor_2699/
The same instances converted into ready-to-run task directories (+ the training index)
Both views cover the identical 2,699 instance_ids. The… See the full description on the dataset page: https://huggingface.co/datasets/Lego-X/Lego-RL-2699.Nemotron-RL-Agentic-Terminal-Pivot-v1
Dataset Description
The Nemotron-RL-Agentic-Terminal-Pivot-v1 dataset provides training samples for reinforcement learning of command-line ("terminal use") LLM agents with the terminus_judge environment in NeMo Gym.
Each record is a single agent decision point extracted from a successful agent trajectory on a terminal task:
responses_create_params.input — the prompt: the task instruction plus the terminal interaction history (prior agent actions and terminal outputs) up to the… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Nemotron-RL-Agentic-Terminal-Pivot-v1.Nemotron-RL-Agentic-Indirect-Prompt-Injection-v1
Nemotron-RL-Agentic-Indirect-Prompt-Injection-v1
Dataset Description:
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 Injection (IPI) attacks hidden inside tool-returned environment data. In each record, the agent receives a benign user request that requires calling a read tool whose output contains an adversarial instruction disguised as legitimate domain content… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Nemotron-RL-Agentic-Indirect-Prompt-Injection-v1.Nemotron-RL-Ultra-Training-Blends
Dataset Description:
This dataset provides Reinforcement Learning (RL) and Multi-teacher On-Policy Distillation (MOPD) training-data blends used by the public Nemotron-3-Ultra post-training recipe. The blends are consumed by the NeMo RL training recipes through the NeMo Gym agent framework, in which each prompt is paired with an agent/environment that returns a verifiable or judge-based reward. Each subset is a separate blend; see the recipe for how the blends are used.
The… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Nemotron-RL-Ultra-Training-Blends.harbor-swesmith-rl-artifacts
Harbor SWE-Smith 强化学习数据产物
本数据集是 Harbor Qwen 工具调用代码智能体强化学习项目使用的冻结任务集，服务于 GRPO、原生价值模型/GAE PPO、训练过程诊断和统一协议评测。
项目已于 2026 年 8 月 30 日完成 P0 评测并进入阶段性归档。本数据集用于保留实验所依赖的数据切分、任务执行文件和审计信息，不代表新的通用代码能力基准。
数据概况
切分
任务数
训练集
187
验证集
42
测试集
38
合计
267
数据覆盖 89 个上游代码仓库。三个切分之间同时执行任务标识和仓库级隔离检查。
正式数据集名称：
swesmith-curated-grpo-267-v1
冻结切分的语义摘要：
ae5df9a3f4a3fc8af44fac420b36529e283839e1bd3de9daba65d5bcda51447d
该值来自 split-manifest.json 的 sha256 字段，用于标识切分语义，不等同于该文件本身的字节级… See the full description on the dataset page: https://huggingface.co/datasets/keryszhan/harbor-swesmith-rl-artifacts.rlvr-reward-hacking-scale-no-conftest-20260909
Matched no-conftest RLVR study 20260909
Complete immutable training, monitoring and comparison trajectories for six models.
All valid outcomes are retained, including refusals, failures and truncations.
The train split name is a dataset-loader convention; record_type identifies
whether a record is training, monitoring, comparison, or a derived judgment.
import json
from datasets import load_dataset
rows = load_dataset("lucabaroni/rlvr-reward-hacking-scale-no-conftest-20260909"… See the full description on the dataset page: https://huggingface.co/datasets/lucabaroni/rlvr-reward-hacking-scale-no-conftest-20260909.Nemotron-RL-Math-v2
Nemotron-RL-Math-v2
Dataset Description:
Nemotron-RL-Math-v2 is a small curated set of mathematical problems selected for reinforcement learning. The dataset is designed for RL training workflows where problems have verifiable answers or other validation signals suitable for Reinforcement Learning from Verifiable Rewards (RLVR).
Problems are sourced from AoPS, StackExchange-derived math data held out from the Nemotron-SFT-Math-v4 SFT set… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Nemotron-RL-Math-v2.Nemotron-RL-ARC-AGI-v1
Dataset Description:
Nemotron-RL-ARC-AGI-v1 is a reinforcement-learning (RL) gym environment dataset of single-step ARC-AGI puzzle prompts intended for RL post-training of large language models. Each row corresponds to one ARC puzzle (a set of (input grid, output grid) demonstration pairs plus a single test input grid) rendered as a text prompt; reward is binary (1.0 / 0.0) determined by exact-match comparison against the ground-truth output grid. No LLM judge is used, no… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Nemotron-RL-ARC-AGI-v1.Nemotron-RL-Science-v1
Dataset Description:
Nemotron-RL-Science-v1 is a reinforcement learning (RL) dataset for science reasoning. Each example provides a problem, a reference answer, and a verifiable RL environment configuration (the agent prompt, the agent/verifier reference, and the answer-extraction template) so that a policy model can be trained with verifiable rewards. It covers three domains (Physics, Biology, and Chemistry), the open-question (OpenQ) format, and two generation setups:… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Nemotron-RL-Science-v1.Nemotron-Math-Proofs-v3-RL
Nemotron-Math-Proofs-v3-RL
Dataset Description:
Nemotron-Math-Proofs-v3-RL is a long-form mathematical reasoning dataset for reinforcement learning. The release contains 9,597 proof-generation prompts.
The dataset uses NeMo Gym-compatible, single-turn user prompts derived from hard proof problems in the AoPS subset of nvidia/Nemotron-Math-Proofs-v1. The train split asks the policy to produce a rigorous solution and self-evaluation. Policy responses and realized… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Nemotron-Math-Proofs-v3-RL.Nemotron-RL-Multichallenge-v1Nemotron-RL-Multichallenge-v1
License: cc-by-4.0
Task Categories: reinforcement-learning, text-generation
Tags: instruction-following, multi-turn, rlvr, nemo-gym
Configs: advanced train split at data/advanced.jsonl; vanilla train split at data/vanilla.jsonl
Domain: instruction following, multi-turn conversation
Modality: text
Capability Breakdown: Multi-turn instruction following [100%]
Source: Hybrid: Manually Collected, Synthetic
Size Bin: <10K
Associated Model Release: Nemotron Ultra… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Nemotron-RL-Multichallenge-v1.Nemotron-RL-InverseIFEval-v1Nemotron-RL-InverseIFEval-v1
License: cc-by-4.0
Task Categories: reinforcement-learning, text-generation
Tags: instruction-following, adversarial, rlvr, nemo-gym
Configs: default train split at data/train.jsonl
Domain: instruction following, adversarial instruction following
Modality: text
Capability Breakdown: Counter-conventional instruction following [100%]
Source: Hybrid: Manually Collected, Synthetic
Size Bin: <10K
Associated Model Release: Nemotron Ultra
Dataset… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Nemotron-RL-InverseIFEval-v1.Nemotron-RL-CFBench-v1Nemotron-RL-CFBench-v1
License: cc-by-4.0
Language: en, ar, hi, zh, ja, ko
Task Categories: reinforcement-learning, text-generation
Tags: instruction-following, constraint-following, rlvr, nemo-gym
Configs: default train split at data/train.jsonl
Domain: instruction following, constraint following
Modality: text
Capability Breakdown: Constraint following [100%]
Source: Hybrid: Manually Collected, Synthetic
Size Bin: <10K
Associated Model Release: Nemotron Ultra
Dataset… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Nemotron-RL-CFBench-v1.Lego-RL-SWE-Bench-Verified
Lego-RL-SWE-Bench-Verified
The 500 SWE-bench Verified instances as ready-to-run harbor RL environments —
the exact evaluation set behind every SWE-bench Verified number in
LEGO-RL, packaged the same way as the training
set Lego-X/Lego-RL-2699 so
one trainer reads both.
Two parallel views of the same 500 instances:
View
Path
What it is
Official SWE-bench records
swebench_verified_official_500/
The upstream princeton-nlp/SWE-bench_Verified rows, verbatim
Harbor RL… See the full description on the dataset page: https://huggingface.co/datasets/Lego-X/Lego-RL-SWE-Bench-Verified.Nemotron-RL-Instruction-Following-Free-Form-Formatting-v1
Dataset Description:
Teaches the model to follow arbitrary text formatting instructions (bullet styles, numbering, delimiters, heading formats, inline emphasis, web-answer structure, etc.) for targeted chat behaviors. Uses explicit Regex and string matching for the reward signal.
This dataset is ready for commercial or non-commercial uses.
Dataset Owner(s):
NVIDIA Corporation
Dataset Creation Date:
Created on: April 10, 2026
Last Modified on: April… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Nemotron-RL-Instruction-Following-Free-Form-Formatting-v1.Nemotron-RL-Safety-v1
Dataset Description:
The Nemotron-RL-Safety-v1 data is designed to provide labeled comparisons necessary to train Reward Models to distinguish between safe, helpful responses and undesired, non-compliant outputs. This dataset is a collection of:
A hybrid (open-source and synthetically generated) collection of prompts designed to elicit different model vulnerabilities, and
Safety Preference pairs: Each prompt is associated with a chosen and rejected response to provide a clear… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Nemotron-RL-Safety-v1.Nemotron-RL-QA-Abstention-v1Nemotron-RL-QA-Abstention-v1
License: cc-by-4.0
Language: en
Task Categories: reinforcement-learning, question-answering, text-generation
Tags: abstention, question-answering, hotpotqa, software-engineering, health, law, rl, rlvr
Configs: default train split at data/train.jsonl
Domain: multi-domain question answering, abstention
Modality: text
Capability Breakdown: Abstention-aware factoid question answering [100%]
Source: Hybrid: Automated, Manually Collected, Synthetic
Size Bin: <10K… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Nemotron-RL-QA-Abstention-v1.Nemotron-RLHF-GenRM-v1
Dataset Description:
This dataset is designed to train Generative Reward Models (GenRMs). It leverages reinforcement learning at scale to train accurate and robust GenRMs that generalize better than traditional Bradley-Terry models and reduce the risk of reward hacking.
The dataset is composed of:
Preference data focused on diverse domains
A synthetic safety blend
The data follows a "meta-prompt" structure where the model is instructed to act as an expert evaluation judge. For… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Nemotron-RLHF-GenRM-v1.social-reasoning-rlhf
Dataset Summary
This repository provides access to a social reasoning dataset that aims to provide signal to how humans navigate social situations, how they reason about them and how they understand each other. It contains questions probing people's thinking and understanding of various social situations.
This dataset was created by collating a set of questions within the following social reasoning tasks:
understanding of emotions
intent recognition
social norms
social… See the full description on the dataset page: https://huggingface.co/datasets/ProlificAI/social-reasoning-rlhf.EnvFactory-RL
EnvFactory: Scaling Tool-Use Agents via Executable Environments Synthesis and Robust RL
## Overview
EnvFactory-RL is a reinforcement learning dataset containing 3,090 tool-use trajectories synthesized using the EnvFactory framework. This dataset is designed for training tool-use agents via Agentic Reinforcement Learning (Agentic RL).
The dataset contains multi-turn tool-use trajectories with implicit human reasoning, generated through… See the full description on the dataset page: https://huggingface.co/datasets/LARK-Lab/EnvFactory-RL.Nemotron-RL-Instruction-Following-Citation-Formatting-v1
Dataset Description:
Teaches the model to cite specific document parts using reference markers like [ref:1], ref:3, etc. Supports single-reference, multi-reference, and inline citations.
This dataset is ready for commercial/non-commercial uses.
Dataset Owner(s):
NVIDIA Corporation
Dataset Creation Date:
Created on: April 10, 2026
Last Modified on: April 10, 2026
Version:
Nemotron-RL-Instruction-Following-CitationFormatting-v1… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Nemotron-RL-Instruction-Following-Citation-Formatting-v1.UltraData-RL-2609
UltraData-RL-2609
📦 UltraData Collection |
🌐 UltraData |
🤗 MiniCPM5 Series
English |
中文
📚 Introduction
UltraData-RL-2609 is the L3 refined data for reinforcement learning within UltraData's L0-L4 tiered data management framework. Built for the RL stage of MiniCPM5-2B post-training, it complements UltraData-SFT-2605 with verifiable-reward tasks. It is also the training corpus used by JustRL II (Scaling Small LLMs to 128K Reasoning with a Critic)… See the full description on the dataset page: https://huggingface.co/datasets/kwakuobeng/UltraData-RL-2609.Nemotron-Cascade-RL-Math
Nemotron-Cascade-RL-Math
Nemotron-Cascade-RL-Math is a diverse and high-quality dataset focused on math reasoning. It serves as the Math RL data for Nemotron-Cascade.
Nemotron-Cascade-RL-MATH contains 14,476 math problems and short answers, covering the data sources from OpenMathReasoning, NuminaMath-CoT, DeepScaleR, AceReason-Math. We conduct data decontamination and filter the sample that has a 9-gram overlap with any test sample in our math benchmarks.
The following are… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Nemotron-Cascade-RL-Math.SPADE-Grounding-Corpus-ToolUse-15K
SPADE grounding corpus: tool use (15k)
Reference documents the SPADE Environment Designer is grounded on when generating multi-turn tool-use environments. 15,552 source files drawn from nvidia/Nemotron-Pretraining-Code-v3.
Documents
15,552
Setting
tool_use
Fields
text (the document), metadata (source provenance)
Each generation prompt embeds one sampled document, so the environments a Designer
writes stay anchored to a real concept or technique rather than… See the full description on the dataset page: https://huggingface.co/datasets/spade-rl/SPADE-Grounding-Corpus-ToolUse-15K.SPADE-Grounding-Corpus-Games-15K
SPADE grounding corpus — games (15k)
Reference documents the SPADE proposer is grounded on when generating cognitive-skill
game environments. 15,000 documents: 10k drawn from a mathematics corpus and 5k from a
science corpus.
Documents
15,000
Setting
games
Fields
Field
Description
text
The document, exactly as embedded in the generation prompt
metadata
domain (mathematics / science) and url (source provenance)
Each generation… See the full description on the dataset page: https://huggingface.co/datasets/spade-rl/SPADE-Grounding-Corpus-Games-15K.rlvr-reward-hacking-transcripts
RLVR reward-hacking full trajectories
This release contains 900 full held-out trajectories from three policies trained with
reinforcement learning from verifiable rewards (RLVR) in a deliberately vulnerable
CodeContests evaluator: 300 each from the final Qwen3.5-9B, GPT-OSS-120B, and Nemotron-3-Super-120B-A12B
checkpoints. Each row preserves the task, tests, complete prompts, native
reasoning, final answer, rendered and sampled token IDs, token log-probabilities, sampling… See the full description on the dataset page: https://huggingface.co/datasets/lucabaroni/rlvr-reward-hacking-transcripts.Nemotron-RL-Jailbreak-Robustness-v1
Dataset Description:
The Nemotron-RL-Jailbreak-Robustness-v1 data is designed to (1) strengthen model robustness against a variety of adversarial jailbreak techniques and (2) at the same time improve adherence to behavioral policies.
This dataset is a collection of hybrid (open-source and synthetically generated) collection of adversarial prompts designed to elicit undesirable behavior from large language models. That's it, just prompts, responses are generated during training… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Nemotron-RL-Jailbreak-Robustness-v1.Ring-lite-rl-data
🤗 Hugging Face
🤖 ModelScope
🖥️ GitHub
Ring-lite-rl-data
This dataset is a curated subset of high-quality problems across mathematics and code domains designed for reinforcement learning in the Ring-lite model. This dataset contains:
Mathematics: Over 39,000 rigorously curated problems sourced from:
Open-source datasets (BigMath, DeepScaleR, DAPO, DeepMath-103K)
Art of Problem Solving (AoPS) contest collections
Code: Approximately 8,400… See the full description on the dataset page: https://huggingface.co/datasets/inclusionAI/Ring-lite-rl-data.HH-RLHF-MetaAligner-Data
Dataset Card for HH-RLHF-MetaAligner-Data
Dataset Summary
This repository provides access to 2 subsets of the MetaAligner data built from HH-RLHF:
The equal-preference subset is included in: equal_train.json and equal_val.json.
The preference subset is included in: preference_train.json, preference_val.json, and preference_test.json.
Disclaimer: As the dialogue data comes from HH-RLHF, the data contain content that may be offensive or upsetting. Topics include, but are… See the full description on the dataset page: https://huggingface.co/datasets/MetaAligner/HH-RLHF-MetaAligner-Data.