datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
Staleness-GRPO-DAPO-Math-17k
Staleness GRPO DAPO Math 17k
The exact 17,005-row training dataset shared by the staleness-cap-2 Qwen2.5-Math-1.5B, Qwen2.5-3B, and Qwen2.5-Math-7B checkpoints, and the staleness-cap-4 Qwen2.5-Math-1.5B checkpoint. All four training manifests record the same SHA-256 for the training file.
Source and processing
Derived from the all configuration of open-r1/DAPO-Math-17k-Processed, itself processed from BytedTsinghua-SIA/DAPO-Math-17k. Source revision:… See the full description on the dataset page: https://huggingface.co/datasets/zbeeb/Staleness-GRPO-DAPO-Math-17k.Manim-grpo-dataset-200
Manim GRPO Dataset 200
200+ cleaned ManimGL scene excerpts and populated metadata bundles for GRPO / reward-model training on mathematical animation code. Each problem is a directory data/problems/MB-XXX/ containing reference.py extracted from 3b1b/videos (years 2022–2026), complete with problem.json, visual_events.json, coverage.json, version_notes.json, and ref_embeddings.npy.
Dataset structure
data/
problems/
MB-001/ … MB-200/
reference.py… See the full description on the dataset page: https://huggingface.co/datasets/nabin2004/Manim-grpo-dataset-200.Go-GRPO-1K
Go-GRPO-1K
Paper | Code
Project Context
The LoGos model uses this dataset to transfer reasoning capabilities acquired from long CoT data to Go tasks. Through mixed fine-tuning and reinforcement learning, the model learns to perform analysis, reasoning, and summarization to select optimal moves on the Go board.
Citation
If you find this dataset useful for your research, please cite:
@misc{ma2026mixingexpertknowledgebring,
title={Mixing Expert Knowledge:… See the full description on the dataset page: https://huggingface.co/datasets/YichuanMa/Go-GRPO-1K.funding-extraction-artifact-data-mix-grpo-mixed-reward
Funding Extraction Training Data
Training, evaluation, and test data for fine-tuning LLMs to extract structured funding metadata (funder names, award IDs, funding schemes, award titles) from academic paper funding statements.
Dataset Structure
data/
├── full/ # Complete unsplit dataset
│ ├── train.jsonl # 5,264 real Crossref funding statements
│ └── synthetic.jsonl # 10,124 LLM-generated funding statements
├── sft/… See the full description on the dataset page: https://huggingface.co/datasets/cometadata/funding-extraction-artifact-data-mix-grpo-mixed-reward.bash-agent-grpo-pairs
Bash Agent GRPO Pairs
Single-turn (intent → shell command) pairs for training a small, local, Claude-Code-style
bash agent with SFT or GRPO. Each record pairs a natural-language objective with exactly
one verifiable bash command, framed as a single-tool bash(command, description) call.
The dataset is designed to be rewardable: the ground-truth command is a deterministic target,
so a shell-equivalence reward (canonical program + flag set + argument comparison) can score… See the full description on the dataset page: https://huggingface.co/datasets/adeelahmad/bash-agent-grpo-pairs.JOSIE-Zero-4-GRPO-Outputs-105
JOSIE-Zero-4: Out-of-Training GRPO Evaluation Generations
This dataset contains 105 model-generated reasoning traces from JOSIE-Zero-4. The samples were selected from a 500-example evaluation on the Math branch of openbmb/UltraData-RL-2609.
These problems were not used to train JOSIE-Zero-4 with GRPO. The evaluation was designed to examine whether reasoning behavior learned through GRPO transfers to data outside the model's training distribution.
Of the 500 evaluated generations… See the full description on the dataset page: https://huggingface.co/datasets/Goekdeniz-Guelmez/JOSIE-Zero-4-GRPO-Outputs-105.Codeforces-Cleaned-GRPO
Codeforces-Cleaned-GRPO
Deep-cleaned for GRPO/RL training | 500 examples | 30 bugs fixed
📋 Dataset Description
Competitive programming problems from Codeforces via Open-R1. This cleaned version removes HTML math notation tags and whitespace issues that would break problem descriptions during training.
Original source: open-r1/codeforces by Open-R1
📊 Cleaning Statistics
Metric
Value
Original examples
500
Clean examples
500… See the full description on the dataset page: https://huggingface.co/datasets/Eyght/Codeforces-Cleaned-GRPO.GRPO-Reasoning-Tools-Cleaned
GRPO-Reasoning-Tools-Cleaned
Deep-cleaned for GRPO/RL training | 1,998 examples | 2,608 bugs fixed
📋 Dataset Description
Structured reasoning and tool-use prompts designed for GRPO training. This cleaned version normalizes whitespace, removes XML reasoning tags, and filters non-English content.
Original source: nphearum/grpo-4k-reasoning-tools by Independent
📊 Cleaning Statistics
Metric
Value
Original examples
2,000
Clean… See the full description on the dataset page: https://huggingface.co/datasets/Eyght/GRPO-Reasoning-Tools-Cleaned.manibench-grpo
ManiBench GRPO Reference Scenes
200 cleaned ManimGL scene excerpts for GRPO / reward-model work on math animation code. Each problem is a folder data/problems/MB-XXX/ with a reference.py extracted from 3b1b/videos (years 2022–2026).
This release is reference code only. Prompt, visual-event, coverage, and version-note JSON files are empty placeholders to fill later. CLIP embeddings and raw video are not included.
Not in this set: the 12 ManiBench pilot / benchmark videos… See the full description on the dataset page: https://huggingface.co/datasets/nabin2004/manibench-grpo.DeepSeek-Reasoning-Cleaned-GRPO
DeepSeek-Reasoning-Cleaned-GRPO
Deep-cleaned for GRPO/RL training | 500 examples | 41 bugs fixed
📋 Dataset Description
Pure reasoning traces from DeepSeek R1 distilled mathematical reasoning data. This cleaned version removes markdown headers, control characters, and whitespace that would confuse model training.
Original source: open-r1/OpenR1-Math-220k by DeepSeek/Open-R1
📊 Cleaning Statistics
Metric
Value
Original examples
500… See the full description on the dataset page: https://huggingface.co/datasets/Eyght/DeepSeek-Reasoning-Cleaned-GRPO.caliber-extension-gemma4-e2b-grpo-rollouts
CALIBER Extension — Gemma4-E2B GRPO Rollouts
Training rollouts from matched GRPO arms on google/gemma-4-E2B-it
(new-prompt template, non-thinking, full bf16, max completion 1500, 150 steps).
Subsets
subset
arm
τ
prior
rows
mean reward_total
accuracy
full schema
caliber
vanilla CALIBER
0.0
—
1600
2.298
0.514
0.664
mink
Min-K% prior
1.0
mink_0.2
4800
2.506
0.520
0.680
minkpp
Min-K++% prior
1.0
minkpp_0.2
4800
2.637
0.541
0.726
Load:
from datasets… See the full description on the dataset page: https://huggingface.co/datasets/dmnsh/caliber-extension-gemma4-e2b-grpo-rollouts.OpenCodeReasoning-Cleaned-GRPO
OpenCodeReasoning-Cleaned-GRPO
Deep-cleaned for GRPO/RL training | 500 examples | 525 bugs fixed
📋 Dataset Description
Code reasoning and critique prompts extracted from NVIDIA's OpenCodeReasoning-2 dataset. This cleaned version removes formatting artifacts, HTML tags, and whitespace issues that would degrade GRPO/RL training quality.
Original source: nvidia/OpenCodeReasoning-2 by NVIDIA
📊 Cleaning Statistics
Metric
Value
Original… See the full description on the dataset page: https://huggingface.co/datasets/Eyght/OpenCodeReasoning-Cleaned-GRPO.HH-RLHF-Cleaned-GRPO
HH-RLHF-Cleaned-GRPO
Deep-cleaned for GRPO/RL training | 1,847 examples | 187 bugs fixed
📋 Dataset Description
Safety and helpfulness prompts extracted from Anthropic's HH-RLHF dataset. This cleaned version removes 100 duplicate prompts and 87 whitespace issues that would waste training compute on redundant data.
Original source: Anthropic/hh-rlhf by Anthropic
📊 Cleaning Statistics
Metric
Value
Original examples
1,947
Clean… See the full description on the dataset page: https://huggingface.co/datasets/Eyght/HH-RLHF-Cleaned-GRPO.arabic-rag-chat-grpo-5K
Arabic multi-turn RAG conversations — GRPO pool (5,259 conversations)
The reinforcement-learning half of
oddadmix/arabic-rag-chat-30K:
same generator, same validator, same schema, disjoint companies. It exists
so GRPO explores fresh knowledge bases instead of taking a second pass over
material the SFT already memorised.
conversations
turns
companies
this pool
5,259
14,018
309
Company-disjointness is exact and verified: this pool shares zero
company_id values… See the full description on the dataset page: https://huggingface.co/datasets/oddadmix/arabic-rag-chat-grpo-5K.pokemon-showdown-grpo-tutorial
Pokémon Showdown GRPO tutorial dataset
Pre-built GRPO records for the ROCm AI Developer Hub tutorial.
Split
File
Records
demo
data/demo.jsonl
64
train
data/train.jsonl
2048
validate
data/validate.jsonl
32
Use via tutorial notebook Step 12 (load_grpo_tutorial_records) or regenerate with prepare_grpo_tutorial_data.py.
Companion scripts: https://github.com/GoldenGrapeGentleman/pokemon-showdown-agent-scripts
battle-game-grpo-tutorial
turn-based battle game GRPO tutorial dataset
Pre-built GRPO records for the ROCm AI Developer Hub tutorial.
Split
File
Records
demo
data/demo.jsonl
64
train
data/train.jsonl
2048
validate
data/validate.jsonl
32
Use via tutorial notebook Step 12 (load_grpo_tutorial_records) or regenerate with prepare_grpo_tutorial_data.py.
Companion scripts: https://github.com/GoldenGrapeGentleman/battle game-showdown-agent-scripts
grpo-oumi-synthetic-document-claims
Dataset Card for GRPO Oumi ANLI Subset
Dataset
This dataset is a reformatted version of the oumi-ai/oumi-synthetic-document-claims dataset, specifically structured for use with the GRPO trainer.
You can find more detailed information about the original dataset at the provided link.
Link: https://huggingface.co/datasets/oumi-ai/oumi-synthetic-document-claims
Dataset Structure
The dataset consists of a list of dictionaries, where each dictionary represents a… See the full description on the dataset page: https://huggingface.co/datasets/Teen-Different/grpo-oumi-synthetic-document-claims.simple-python-grpo
simple-python-grpo
A curated set of simple Python function problems for GRPO / RLVR fine-tuning.
Each row has a natural-language description, a function signature, and 3
auto-verified test assertions (generated by running a reference implementation,
so every test is correct by construction). The reference is NOT included — the
model must generate the body and is rewarded when the tests pass.
Fields: name, prompt, func_prompt, tests (newline-separated asserts),
setup_code.
Built… See the full description on the dataset page: https://huggingface.co/datasets/sagecodes/simple-python-grpo.wisconsin-building-codes-grpo
Wisconsin Building Codes Q&A Dataset (GRPO-Formatted)
This dataset is a version of the Wisconsin Building Codes Q&A Dataset formatted specifically for Grouped-Reward-Optimization (GRPO) training with libraries like TRL and unsloth.
Dataset Description
This dataset contains 13,200 prompts designed for training preference models. Each record includes a user prompt (prompt), a "chosen" high-quality response, and a placeholder for a "rejected" response.
Training samples: 11… See the full description on the dataset page: https://huggingface.co/datasets/carlscape/wisconsin-building-codes-grpo.DATA_FOR_GRPO_ZERO
DATA_FOR_GRPO_ZERO
GRPO-ZEROおよび、そのSFT（教師あり微調整）検証を目的として作成した日本語データセットです。多様なタスク形式を一つのコーパスにまとめ、学習・評価の両方で扱いやすい構成にしています。
概要
形式: JSON Lines（.jsonl）
言語: 日本語
ライセンス: ODC-By（Open Data Commons Attribution License）
生成: DeepSeek-AI公式APIで利用可能な DeepSeek-V3-0324 による合成データ
収録タスクと件数
本データは以下の内訳で構成されています（合計 1,500件）。
区分
件数
説明
簡単な質問
300
一般的な短文QA・基礎的な問い合わせ
中難易度な質問
300
条件付きの質問、文脈を踏まえた回答など
算数/初等数学
100
計算・文章題・初等的推論
RP（ロールプレイ）タスク
250
役割設定に基づく対話・応答
小説生成タスク
250… See the full description on the dataset page: https://huggingface.co/datasets/TeamDelta/DATA_FOR_GRPO_ZERO.grpo-oumi-c2d-d2c-subset
Dataset Card for GRPO Oumi ANLI Subset
Dataset
This dataset is a reformatted version of the oumi-ai/oumi-c2d-d2c-subset dataset, specifically structured for use with the GRPO trainer.
You can find more detailed information about the original dataset at the provided link.
Link: https://huggingface.co/datasets/oumi-ai/oumi-c2d-d2c-subset
Dataset Structure
The dataset consists of a list of dictionaries, where each dictionary represents a single data instance with… See the full description on the dataset page: https://huggingface.co/datasets/Teen-Different/grpo-oumi-c2d-d2c-subset.grpo-opencoder-mini
grpo-opencoder-mini
This dataset was created by converting another dataset to JSONL format.
Files
dataset.jsonl: Dataset in JSONL format
Usage
from datasets import load_dataset
dataset = load_dataset("Nutanix/grpo-opencoder-mini", data_files="dataset.jsonl")
grpo-oumi-synthetic-claims
Dataset Card for GRPO Oumi ANLI Subset
Dataset
This dataset is a reformatted version of the TEEN-D/grpo-oumi-anli-subset dataset, specifically structured for use with the GRPO trainer.
You can find more detailed information about the original dataset at the provided link.
Link: https://huggingface.co/datasets/oumi-ai/oumi-synthetic-claims
Dataset Structure
The dataset consists of a list of dictionaries, where each dictionary represents a single data instance… See the full description on the dataset page: https://huggingface.co/datasets/Teen-Different/grpo-oumi-synthetic-claims.deepseek_grpo_correct_6144
DeepSeek GRPO Correct 6144
Filtered GRPO training subset generated from deepseek-reasoner math generations.
Files
train.jsonl: filtered training examples with prompt, solution, dataset_index, and DeepSeek metadata.
metadata.json: filtering metadata.
Filter
Rows are kept when the raw generation is successful, stopped, correct, deduplicated by dataset_index, and has usage_total_tokens <= 6144.
Summary
Rows: 7576
Max total tokens: 6144
Source raw… See the full description on the dataset page: https://huggingface.co/datasets/igreck/deepseek_grpo_correct_6144.deepseek_grpo_correct_8192
DeepSeek GRPO Correct 8192
Filtered GRPO training subset generated from deepseek-reasoner math generations.
Files
train.jsonl: filtered training examples with prompt, solution, dataset_index, and DeepSeek metadata.
metadata.json: filtering metadata.
Filter
Rows are kept when the raw generation is successful, stopped, correct, deduplicated by dataset_index, and has usage_total_tokens <= 8192.
Summary
Rows: 9002
Max total tokens: 8192
Source raw… See the full description on the dataset page: https://huggingface.co/datasets/igreck/deepseek_grpo_correct_8192.grpo-opencoder-small
grpo-opencoder-small
This dataset was created by converting another dataset to JSONL format.
Files
dataset.jsonl: Dataset in JSONL format
Usage
from datasets import load_dataset
dataset = load_dataset("Nutanix/grpo-opencoder-small", data_files="dataset.jsonl")
grpo-oumi-anli-subset
Dataset Card for GRPO Oumi ANLI Subset
Dataset
This dataset is a reformatted version of the oumi-ai/oumi-anli-subset dataset, specifically structured for use with the GRPO trainer.
You can find more detailed information about the original dataset at the provided link.
Link: https://huggingface.co/datasets/oumi-ai/oumi-anli-subset
Dataset Structure
The dataset consists of a list of dictionaries, where each dictionary represents a single data instance with a… See the full description on the dataset page: https://huggingface.co/datasets/Teen-Different/grpo-oumi-anli-subset.grpo-opencoder-50k
grpo-opencoder-50k
This dataset was created by converting another dataset to JSONL format.
Files
dataset.jsonl: Dataset in JSONL format
Usage
from datasets import load_dataset
dataset = load_dataset("Nutanix/grpo-opencoder-50k", data_files="dataset.jsonl")
deepseek_grpo_correct_4096
DeepSeek GRPO Correct 4096
Filtered GRPO training subset generated from deepseek-reasoner math generations.
Files
train.jsonl: filtered training examples with prompt, solution, dataset_index, and DeepSeek metadata.
metadata.json: filtering metadata.
Filter
Rows are kept when the raw generation is successful, stopped, correct, deduplicated by dataset_index, and has usage_total_tokens <= 4096.
Summary
Rows: 5502
Max total tokens: 4096
Source raw… See the full description on the dataset page: https://huggingface.co/datasets/igreck/deepseek_grpo_correct_4096.deepseek_grpo_correct_2048
DeepSeek GRPO Correct 2048
Filtered GRPO training subset generated from deepseek-reasoner math generations.
Files
train.jsonl: filtered training examples with prompt, solution, dataset_index, and DeepSeek metadata.
metadata.json: filtering metadata.
Filter
Rows are kept when the raw generation is successful, stopped, correct, deduplicated by dataset_index, and has usage_total_tokens <= 2048.
Summary
Rows: 2153
Max total tokens: 2048
Source raw… See the full description on the dataset page: https://huggingface.co/datasets/igreck/deepseek_grpo_correct_2048.