datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
Math-RL-Tasks
Ulam AI Math RL Tasks
Forty original, verifier-backed mathematical reasoning tasks packaged as ten
independent RL environments. The collection spans advanced graduate exercises,
research-style exact computation and structural generalization problems in
algebraic geometry, arithmetic geometry, combinatorics, topology, probability
and spectral analysis.
Each suite pairs a runnable rl_env/ with a preserved blind_run/ by
GPT-5.6 Sol Pro. The model name describes the evaluation actor… See the full description on the dataset page: https://huggingface.co/datasets/ulamai/Math-RL-Tasks.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-litmus-bench-v0.1
Dataset Description:
Litmus-Bench v0.1 is an open dataset for training and evaluating chemical reasoning in language models. It includes 5,232 training questions and 482 test questions, each in short-answer format and was created from the ChEMBL dataset with RDKit descriptors requiring short answers. The dataset is for RL training.
This dataset is released as part of NVIDIA NeMo-Gym, an open-source library within the NVIDIA NeMo framework, designed for large-scale, verifiable… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Nemotron-RL-litmus-bench-v0.1.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.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.Open-RL
Open-RL
Dataset Summary
This dataset contains self-contained, verifiable, and unambiguous STEM reasoning problems across Physics, Mathematics, Biology, and Chemistry.
Each problem:
Requires multi-step reasoning
Involves symbolic manipulation and/or numerical computation
Has a deterministic, objectively verifiable final answer
The problems were evaluated against contemporary large language models. Observed pass rates indicate that the tasks are non-trivial yet… See the full description on the dataset page: https://huggingface.co/datasets/TuringEnterprises/Open-RL.selective-readtwice-14b-rl-data-20260925
Actual 14B selective RL training data — 2026-09-25 cutoff
This repository contains the actual retained dataset used for selective RL following the completed 14B SFT checkpoint. The user explicitly requested stopping unfinished outline generation and starting RL with completed data. It is not the originally planned full 19,456-view dataset.
Source length
Retained training views
Short (32K bucket)
16,384
56K
1,024
112K
1,024
224K
14
Total
18,446
There are… See the full description on the dataset page: https://huggingface.co/datasets/Xirui1208/selective-readtwice-14b-rl-data-20260925.RLEVThis repository contains the QA data presented in the paper Every Question Has Its Own Value: Reinforcement Learning with Explicit Human Values.
We propose Reinforcement Learning with Explicit Human Values (RLEV), a method that aligns Large Language Model (LLM) optimization directly with quantifiable human value signals. While Reinforcement Learning with Verifiable Rewards (RLVR) effectively
trains models in objective domains using binary correctness rewards, it overlooks that not all tasks… See the full description on the dataset page: https://huggingface.co/datasets/sarosavo/RLEV.supersede-rl-episodes
Supersede RL Episodes
Procedurally-generated supersession episodes for reinforcement learning — the training data
behind vedant33/supersede-qwen2.5-3b-grpo-lora
and the Supersede environment.
Paper
Code
Environment
Model
arXiv · DOI
GitHub
Prime Intellect Hub
vedant33/supersede-qwen2.5-3b-grpo-lora
What this is
Each episode is a short multi-session conversation in which a fact about the user changes one or
more times (they move city, switch jobs… See the full description on the dataset page: https://huggingface.co/datasets/vedant33/supersede-rl-episodes.Variational-DAPO
Dataset Card for SvS/Variational-DAPO
[🌐 Website] •
[🤗 Dataset] •
[📜 Paper] •
[🐱 GitHub] •
[🐦 Twitter] •
[📕 Rednote]
This dataset consists of 314k variational problems synthesized by the Qwen2.5-32B-Instruct policy during RLVR training on DAPO-17k using the SvS strategy for 600-step training, each accompanied by reference answers.The variational problems undergo a min_hash deduplication with a threshold of 0.85.
Data Loading
from datasets import… See the full description on the dataset page: https://huggingface.co/datasets/RLVR-SvS/Variational-DAPO.Phy-RL
IPhO Physics RLVR
English IPhO physics problems curated into RLVR-ready question/answer rows.
Each admitted row contains:
problem_text: full prompt context plus the focused question
shared_context: reusable context needed to answer the question
question: focused answerable question
official_solution: solution evidence supporting the answer
answers: structured verifier targets with value, unit, answer type, tolerance, verifier, equivalent forms, and subproblem id
split: train or… See the full description on the dataset page: https://huggingface.co/datasets/Darkyy/Phy-RL.RL-preview
QIMMA-oriented Arabic MCQ dataset v1.0
This frozen release contains 477 training-ready Arabic multiple-choice
questions: 81 from the reviewed v2 pilot and 396 newly
accepted questions generated with openai/gpt-oss-120b.
Use
Use training_ready.jsonl as the canonical merged training pool. Each record contains
the question, four choices, answer, explanation, source attribution, difficulty,
quality adjudication, and release origin. batch_training_ready.jsonl contains… See the full description on the dataset page: https://huggingface.co/datasets/Mohaddz/RL-preview.lemonseed-rl-tasks-v2
lemonseed-rl-tasks-v2
LemonSeed — GRPO RL tasks v2 (vocab MC / antonym MC / cloze MC / grammar / dialogue / arithmetic).
Contents
rl_tasks_v2.jsonl (12492 rows)
Format
JSON Lines (.jsonl), one example per line.
Provenance
Synthetic, generated programmatically for the LemonSeed 1.5B project (by Geramy L. Loveless). Data authored by Michael Anthony Falabella.
Vietnamese-beyond-rlhf-reward-single-round-gg-translatedNemotron-RL-litmus-bench-v0.1
Dataset Description:
Litmus-Bench v0.1 is an open dataset for training and evaluating chemical reasoning in language models. It includes 5,232 training questions and 482 test questions, each in short-answer format and was created from the ChEMBL dataset with RDKit descriptors requiring short answers. The dataset is for RL training.
This dataset is released as part of NVIDIA NeMo-Gym, an open-source library within the NVIDIA NeMo framework, designed for large-scale, verifiable… See the full description on the dataset page: https://huggingface.co/datasets/Arsh9210/Nemotron-RL-litmus-bench-v0.1.lemonseed-rl-tasks-cogen-snapshot
lemonseed-rl-tasks-cogen-snapshot
LemonSeed — RL co-gen task snapshot.
Contents
rl_tasks_cogen_snapshot.jsonl (1809 rows)
Format
JSON Lines (.jsonl), one example per line.
Provenance
Synthetic, generated programmatically for the LemonSeed 1.5B project (by Geramy L. Loveless). Data authored by Michael Anthony Falabella.
GLM-5.1-Reasoning-1M-Cleaned
GLM-5.1-Reasoning-1M-Cleaned
GLM-5.1-Reasoning-1M-Cleaned is a cleaned and reformatted derivative of Kassadin88/GLM-5.1-1000000x. It preserves the original four-subset layout (main, PHD-Science, Multilingual-STEM, Math) while converting every example into a unified SFT-ready schema with explicit conversations, input, output, domain, and meta fields.
This release was prepared from the original dataset published by Kassadin88.
Summary
Teacher model in the data: GLM-5.1… See the full description on the dataset page: https://huggingface.co/datasets/rlandismd/GLM-5.1-Reasoning-1M-Cleaned.SimScholar-RL
S3 RL Dataset
Synthetic literature-search questions for agentic reinforcement learning.
Code ·
S3 collection ·
Source corpus
The S3 RL dataset contains 13,000 synthetic, single-hop questions for training
and analyzing scientific-literature search agents. Each question is grounded
in one paper from a fixed corpus of approximately 1.12 million
computer-science papers.
At a glance
Property
Value
Rows
13,000
Language
English
Task
Tool-assisted… See the full description on the dataset page: https://huggingface.co/datasets/trillionlabs/SimScholar-RL.lemonseed-rl-tasks-cogen
lemonseed-rl-tasks-cogen
LemonSeed — RL co-generated chat tasks (teacher-authored).
Contents
rl_tasks_cogen.jsonl (2500 rows)
Format
JSON Lines (.jsonl), one example per line.
Provenance
Synthetic, generated programmatically for the LemonSeed 1.5B project (by Geramy L. Loveless). Data authored by Michael Anthony Falabella.
lemonseed-rl-chat-tasks
lemonseed-rl-chat-tasks
LemonSeed — chat-alignment RL tasks (prompt/gold single-turn).
Contents
rl_chat_tasks.jsonl (9852 rows)
Format
JSON Lines (.jsonl), one example per line.
Provenance
Synthetic, generated programmatically for the LemonSeed 1.5B project (by Geramy L. Loveless). Data authored by Michael Anthony Falabella.
Open-RL
Open-RL
Dataset Summary
This dataset contains self-contained, verifiable, and unambiguous STEM reasoning problems across Physics, Mathematics, Biology, and Chemistry.
Each problem:
Requires multi-step reasoning
Involves symbolic manipulation and/or numerical computation
Has a deterministic, objectively verifiable final answer
The problems were evaluated against contemporary large language models. Observed pass rates indicate that the tasks are non-trivial yet… See the full description on the dataset page: https://huggingface.co/datasets/Jcmschott/Open-RL.Open-RL
Open-RL
Dataset Summary
This dataset contains self-contained, verifiable, and unambiguous STEM reasoning problems across Physics, Mathematics, Biology, and Chemistry.
Each problem:
Requires multi-step reasoning
Involves symbolic manipulation and/or numerical computation
Has a deterministic, objectively verifiable final answer
The problems were evaluated against contemporary large language models. Observed pass rates indicate that the tasks are non-trivial yet… See the full description on the dataset page: https://huggingface.co/datasets/Dembo1/Open-RL.Open-RL
Open-RL
Dataset Summary
This dataset contains self-contained, verifiable, and unambiguous STEM reasoning problems across Physics, Mathematics, Biology, and Chemistry.
Each problem:
Requires multi-step reasoning
Involves symbolic manipulation and/or numerical computation
Has a deterministic, objectively verifiable final answer
The problems were evaluated against contemporary large language models. Observed pass rates indicate that the tasks are non-trivial yet… See the full description on the dataset page: https://huggingface.co/datasets/Erikarhart/Open-RL.Nemotron-RL-litmus-bench-v0.1
Dataset Description:
Litmus-Bench v0.1 is an open dataset for training and evaluating chemical reasoning in language models. It includes 5,232 training questions and 482 test questions, each in short-answer format and was created from the ChEMBL dataset with RDKit descriptors requiring short answers. The dataset is for RL training.
This dataset is released as part of NVIDIA NeMo-Gym, an open-source library within the NVIDIA NeMo framework, designed for large-scale, verifiable… See the full description on the dataset page: https://huggingface.co/datasets/Iambackup/Nemotron-RL-litmus-bench-v0.1.nemotron-nano-rl-mcqa-19k
Nemotron Nano RL MCQA 19K
Nemotron Nano RL MCQA 19K is a 19,670-example English multiple-choice question answering dataset prepared for reinforcement learning with verifiable rewards (RLVR). Its nano_v3_sft_profiled_stem_mcqa identifier and schema correspond to the knowledge-MCQA component of NVIDIA's Nemotron-3-Nano-RL-Training-Blend, represented here in a compact prompt / label / metadata JSONL format.
Each record contains a formatted user prompt, the correct option identifier… See the full description on the dataset page: https://huggingface.co/datasets/wflying/nemotron-nano-rl-mcqa-19k.lemonseed-rl-tasks-cogen-pilot
lemonseed-rl-tasks-cogen-pilot
LemonSeed — RL co-gen pilot tasks.
Contents
rl_tasks_cogen_pilot.jsonl (61 rows)
Format
JSON Lines (.jsonl), one example per line.
Provenance
Synthetic, generated programmatically for the LemonSeed 1.5B project (by Geramy L. Loveless). Data authored by Michael Anthony Falabella.
Open-RL
Open-RL
Dataset Summary
This dataset contains self-contained, verifiable, and unambiguous STEM reasoning problems across Physics, Mathematics, Biology, and Chemistry.
Each problem:
Requires multi-step reasoning
Involves symbolic manipulation and/or numerical computation
Has a deterministic, objectively verifiable final answer
The problems were evaluated against contemporary large language models. Observed pass rates indicate that the tasks are non-trivial yet… See the full description on the dataset page: https://huggingface.co/datasets/QiuMing9/Open-RL.lemonseed-rl-tasks
lemonseed-rl-tasks
LemonSeed — GRPO RL tasks (vocab/cloze/antonym/dialogue-QA, machine-checkable).
Contents
rl_tasks.jsonl (9583 rows)
Format
JSON Lines (.jsonl), one example per line.
Provenance
Synthetic, generated programmatically for the LemonSeed 1.5B project (by Geramy L. Loveless). Data authored by Michael Anthony Falabella.
zhihu_3k_rlhf_train
Note
some rm data from public dataset
format
{
"history": [
"query1", "answer1",
"query2", "answer2"
],
"prompt": "query",
"input": "input for query",
"output": [
"output rank1",
"output rank2",
"output rank3"
]
}
Thanks
beyond/rlhf-reward-single-round-trans_chinese :
dikw/hh_rlhf_cn
liyucheng/zhihu_rlhf_3k
rlaif_training_fictional_patriot_experiment
RLAIF Training Data: The "Honest Patriot" Experiment
Dataset Description
This dataset contains 250 synthetic training examples generated using a Constitutional AI (RLAIF) approach.
It was designed to test the ability of Small Language Models (SLMs) to adhere to a complex, conflicting set of behavioral instructions ("The Constitution") that requires balancing extreme politeness, unwavering logical factuality, and patriotic bias toward a fictional country.
The… See the full description on the dataset page: https://huggingface.co/datasets/TitleOS/rlaif_training_fictional_patriot_experiment.