datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
deepscaler-preview-verl
DeepScaleR-Preview VERL
📊 Dataset Summary
This dataset contains 35,789 mathematical reasoning problems in VERL format, processed from agentica-org/DeepScaleR-Preview-Dataset.
Key Features:
35,789 high-quality math problems
Converted to VERL format for reward modeling
Verified ground truth answers
Ready for reinforcement learning training
🔗 Source Dataset
Original Repository
Repository:… See the full description on the dataset page: https://huggingface.co/datasets/sungyub/deepscaler-preview-verl.verl-deepscaler-clean
verl-deepscaler-clean
A cleaned, deduplicated release of the DeepScaleR math problem set, packaged in the parquet layout
that verl expects from its RLHFDataset, so it can be dropped
straight into a GRPO / rule-based-reward run (for example, when re-running an R1-style "aha moment"
experiment on a smaller model).
The whole dataset lives in a single parquet file at the repo root: verl_deepscaler.parquet.
Headline numbers
Total lines read
3038
Records… See the full description on the dataset page: https://huggingface.co/datasets/dusersad12/verl-deepscaler-clean.code-verl-unified
Unified Code VERL Dataset
Overview
This dataset aggregates seven code-reasoning collections into a single VERL-formatted repository containing approximately 958,539 unique problems. The compilation prioritizes consistent extra_info structure across all source materials for seamless compatibility with VERL training frameworks.
Dataset Composition
Seven distinct splits comprise the collection:
Split
Problems
Percentage
Format
kodcode_v1_verl
434,876… See the full description on the dataset page: https://huggingface.co/datasets/sungyub/code-verl-unified.dapo-math-17k-verl
DAPO-Math-17K VERL
📊 Dataset Summary
This dataset contains 17,147 mathematical reasoning problems in VERL format, processed from haizhongzheng/DAPO-Math-17K-cleaned.
Key Features:
17,147 high-quality math problems
Converted to VERL format for reward modeling
Verified ground truth answers
Ready for reinforcement learning training
🔗 Source Dataset
Original Repository
Repository: haizhongzheng/DAPO-Math-17K-cleaned
License: Apache 2.0… See the full description on the dataset page: https://huggingface.co/datasets/sungyub/dapo-math-17k-verl.math-verl-unified
Math-VERL Unified Collection
📊 Dataset Summary
A unified collection of 9 high-quality mathematical reasoning datasets totaling 2,269,166 deduplicated problems, all converted to VERL format for reinforcement learning applications. This collection combines diverse mathematical content from competition-level problems to advanced reasoning tasks.
Key Features:
2.27M deduplicated samples from 9 curated sources
Inter-dataset deduplication applied (v3.0) - 12.7%… See the full description on the dataset page: https://huggingface.co/datasets/sungyub/math-verl-unified.openr1-math-verl
OpenR1-Math VERL
📊 Dataset Summary
This dataset contains 120,387 mathematical reasoning problems in VERL format, processed from open-r1/OpenR1-Math-220k.
Key Features:
120,387 high-quality math problems
Converted to VERL format for reward modeling
Verified ground truth answers
Ready for reinforcement learning training
🔗 Source Dataset
Original Repository
Repository: open-r1/OpenR1-Math-220k
License: Apache 2.0
Authors: Open-R1 Team… See the full description on the dataset page: https://huggingface.co/datasets/sungyub/openr1-math-verl.qa-verl-unified
Unified QA VERL Dataset
A unified collection of 5 high-quality question-answering and reasoning datasets in VERL format, deduplicated and optimized for reinforcement learning training.
Dataset Summary
This dataset combines 5 diverse QA and reasoning datasets into a single unified collection:
Total Problems: 86,379 unique problems (after 0.00% deduplication)
Original Size: 0 problems (before deduplication)
Format: VERL (Volcano Engine Reinforcement Learning)
Language:… See the full description on the dataset page: https://huggingface.co/datasets/sungyub/qa-verl-unified.reasoning-gym-verl-datasets
reasoning-gym-verl-datasets
This dataset contains procedurally generated reasoning tasks from the Reasoning Gym (r-gym) framework, structured and pre-processed in parquet format for training models with veRL.
These datasets were used to train MauroPello/Qwen3-1.7B-RL-final using GRPO (Group Relative Policy Optimization).
Dataset Splits & Structure
Split Name
Path
Size (Examples)
Description
train
train.parquet
100,000
Raw training set containing… See the full description on the dataset page: https://huggingface.co/datasets/MauroPello/reasoning-gym-verl-datasets.verl_deepscaler
DeepScaleR for verl (full cleaned build)
RL-ready dataset in verl parquet format, rebuilt from the
full three-shard DeepScaleR mirror dump.
Build process
Merged the three mirror shards (deepscaler_shard_00.json, deepscaler_shard_01.json,
deepscaler_shard_02.json) in ascending filename order, keeping each shard's record order.
Dropped records whose problem or answer was missing or whitespace-only (78 dropped).
An empty solution is normal in this export and is… See the full description on the dataset page: https://huggingface.co/datasets/dusersad12/verl_deepscaler.bigcodebench_verl
BigCodeBench (verl eval set)
bigcode/bigcodebench v0.1.4
(1137 rows, data_source=bigcodebench) plus
bigcode/bigcodebench-hard v0.1.4
(146 rows, data_source=bigcodebench_hard; a subset of full, duplicated on purpose so
full and hard pass@1 aggregate separately), converted to the verl rule-reward schema.
Built by verl/scripts/data/bigcodebench.py; source license Apache-2.0.
3 tasks dropped (BigCodeBench/101, /590, /1005; 2 of them also in hard):
their canonical solutions fetch… See the full description on the dataset page: https://huggingface.co/datasets/OctoReasoner/bigcodebench_verl.mercury_verl
Mercury (verl efficiency eval set)
The eval split of Elfsong/Mercury
(arXiv 2402.07844; 256 LeetCode-style tasks;
the train split ships no test cases and is not gradable), converted to the verl
rule-reward schema by verl/scripts/data/mercury.py. Source license
CC-BY-NC-4.0 (non-commercial) -- this conversion keeps that license.
Every row's ground truth carries the full official scoring contract: entry
point, the task's convert_offline/evaluate_offline hooks (lctk linked-list /… See the full description on the dataset page: https://huggingface.co/datasets/OctoReasoner/mercury_verl.cyberseceval_verl
CyberSecEval Instruct (verl safety eval set)
The static, prompt->completion Instruct sub-eval of Meta's CyberSecEval
(PurpleLlama, arXiv 2312.04724), converted
to the verl rule-reward schema by verl/scripts/data/cyberseceval.py:
data_source=code_sec_cse_instruct (1681 rows) -- from
CybersecurityBenchmarks/datasets/instruct/instruct-v2.json.
Autocomplete is intentionally not included. CyberSecEval's Autocomplete
sub-eval is the same insecure-snippet corpus as Instruct (1651 of… See the full description on the dataset page: https://huggingface.co/datasets/OctoReasoner/cyberseceval_verl.eurus-2-math-verl
Eurus-2 Math VERL
📊 Dataset Summary
This dataset contains 283,612 mathematical reasoning problems in VERL format, processed from PRIME-RL/Eurus-2-RL-Data.
Key Features:
283,612 high-quality math problems
Converted to VERL format for reward modeling
Verified ground truth answers
Ready for reinforcement learning training
🔗 Source Dataset
Original Repository
Repository: PRIME-RL/Eurus-2-RL-Data
License: MIT
Authors: PRIME-RL Team… See the full description on the dataset page: https://huggingface.co/datasets/sungyub/eurus-2-math-verl.effibench_verl
EffiBench (verl efficiency eval set)
A canonical-validated subset of DONG19/EffiBench
(arXiv 2402.02037; 1000 LeetCode-style
Python tasks), converted to the verl rule-reward schema by
verl/scripts/data/effibench.py. License: unspecified upstream (no license
tag on the HF dataset or the GitHub repo) -- treat as research-only until
clarified.
Row filter (854/1000 kept). The upstream test_case column is broken for a
sizeable minority of rows: tree/linked-list tasks contain literal… See the full description on the dataset page: https://huggingface.co/datasets/OctoReasoner/effibench_verl.ifbench-verl
IFBench-VERL: Instruction Following Evaluation Dataset for VERL Training
Overview
IFBench-VERL is a comprehensive instruction-following evaluation dataset formatted for VERL (Versatile Reinforcement Learning) training pipelines. This dataset contains 95,373 high-quality examples with 54 different constraint types, enabling systematic training and evaluation of instruction-following capabilities in language models.
The dataset is converted from… See the full description on the dataset page: https://huggingface.co/datasets/sungyub/ifbench-verl.big-math-rl-verl
Big-Math-RL VERL
📊 Dataset Summary
This dataset contains 196,329 mathematical reasoning problems in VERL format, processed from SynthLabsAI/Big-Math-RL-Verified.
Key Features:
196,329 high-quality math problems
Converted to VERL format for reward modeling
Verified ground truth answers
Ready for reinforcement learning training
🔗 Source Dataset
Original Repository
Repository: SynthLabsAI/Big-Math-RL-Verified
License: Apache 2.0… See the full description on the dataset page: https://huggingface.co/datasets/sungyub/big-math-rl-verl.verl-aha-moment-dataset
Verl Dataset for DeepSeek-R1 "Aha Moment" Reproduction
This dataset is prepared for training with the Verl framework
to reproduce the "aha moment" phenomenon observed in DeepSeek-R1-Zero.
What is the "Aha Moment"?
During RL training without any SFT, DeepSeek-R1-Zero spontaneously developed:
Self-verification: Checking answers within `` tags
Long chain-of-thought: Extended reasoning traces
Backtracking: "Wait, that seems wrong..." behavior
Metacognition:… See the full description on the dataset page: https://huggingface.co/datasets/cxzsad12e/verl-aha-moment-dataset.orz-math-72k-verl
ORZ-Math-72K VERL
📊 Dataset Summary
This dataset contains 44,812 mathematical reasoning problems in VERL format, processed from vwxyzjn/rlvr_orz_math_72k_collection_extended.
Key Features:
44,812 high-quality math problems
Converted to VERL format for reward modeling
Verified ground truth answers
Ready for reinforcement learning training
🔗 Source Dataset
Original Repository
Repository: vwxyzjn/rlvr_orz_math_72k_collection_extended… See the full description on the dataset page: https://huggingface.co/datasets/sungyub/orz-math-72k-verl.verl-deepscaler-curated
verl DeepScaleR Curated
A cleaned, de-duplicated and evaluation-safe training split derived from the
DeepScaleR-Preview-Dataset,
reformatted for rule-based-reward RL post-training with
verl.
Total examples: 38,783 (from 41,705 raw records read across three source batches).
Row format
Each row follows the verl dataset_row template:
field
value
data_source
"DeepScaleR"
prompt
[{"role": "user", "content": <problem text>}]
ability
"math"
reward_model… See the full description on the dataset page: https://huggingface.co/datasets/dusersad12/verl-deepscaler-curated.verl-deepscaler-cleaned
verl DeepScaleR - cleaned & split
A cleaned version of the DeepScaleR-Preview-Dataset
(40,315 math problem/answer pairs used for R1-style "aha moment" reproductions) converted into the
parquet layout that verl expects for rule-based math RL (GRPO / PPO).
Cleaning
stage
entries
raw entries
40,315
dropped: missing / empty answer
6
dropped: duplicate problem (whitespace-insensitive, first occurrence kept)
1,180
final
39,129
Splits… See the full description on the dataset page: https://huggingface.co/datasets/dusersad12/verl-deepscaler-cleaned.verl-code-corpus-track-a-file-split
archit11/verl-code-corpus-track-a-file-split
Repository-specific code corpus extracted from the verl project and split by file for training/evaluation.
What is in this dataset
Source corpus: data/code_corpus_verl
Total files: 214
Train files: 172
Validation files: 21
Test files: 21
File type filter: .py
Split mode: file (file-level holdout)
Each row has:
file_name: flattened source file name
text: full file contents
Training context
This dataset was used… See the full description on the dataset page: https://huggingface.co/datasets/archit11/verl-code-corpus-track-a-file-split.verl-mathmix
verl-mathmix
A curated math QA corpus packaged for RL post-training with the verl
framework, in verl's standard chat-format parquet layout (rule-based verifiable rewards).
Row counts
Training rows: 2187 (verl_mathmix_train.parquet)
Validation rows: 243 (verl_mathmix_val.parquet)
Total: 2430 rows
Schema (per row, exactly as in format.json)
column
value
data_source
DeepScaleR / GSM8K / AMC-AIME (source the kept record came from)
prompt… See the full description on the dataset page: https://huggingface.co/datasets/dusersad12/verl-mathmix.verl-aha-moment-dataset
Verl "Aha Moment" Dataset for DeepSeek-R1 Reproduction
Overview
This dataset is built for use with the Verl framework to reproduce the "aha moment" phenomenon observed in DeepSeek-R1-Zero. It combines two Hugging Face math-reasoning datasets into a single Parquet file conforming to the format specified in format.json.
Composition
Source
Rows
Description
agentica-org/DeepScaleR-Preview-Dataset
40,315
~40K competition math problems (AIME… See the full description on the dataset page: https://huggingface.co/datasets/dsa1dsa12/verl-aha-moment-dataset.skywork-or1-code-verl
Skywork-OR1-RL-Data (Code Split - Cleaned)
This dataset contains 14,057 coding problems from the Skywork-OR1-RL-Data dataset, cleaned of repetitive instruction prefixes. The dataset is in VERL format and includes model difficulty ratings for reinforcement learning applications.
Source: Skywork/Skywork-OR1-RL-Data (Code split only)
License: MIT
Dataset Structure
The dataset follows the VERL format with the following fields:
data_source (string): Dataset source… See the full description on the dataset page: https://huggingface.co/datasets/sungyub/skywork-or1-code-verl.ptdbench-verl-implementation-torch-functional-dataset
PTDBench dataset snapshot: torch_functional
This repository stores the immutable runtime dataset snapshot for one
materialized PTDBench task. It intentionally excludes model weights and
training checkpoints.
PTDBench family: verl_implementation
Source evaluation metric: val-core/taco/acc/mean@1
Provenance: Processed from local TACO EASY (drop picture_num != 0); 8368 train / 184 test rows; bytes identical to task_function_call.
License: Apache-2.0
The artifact manifest records… See the full description on the dataset page: https://huggingface.co/datasets/LIF1014/ptdbench-verl-implementation-torch-functional-dataset.skywork-or1-math-verl
Skywork-OR1-Math VERL
📊 Dataset Summary
This dataset contains 39,202 mathematical reasoning problems in VERL format, processed from Skywork/Skywork-OR1-RL-Data.
Key Features:
39,202 high-quality math problems
Converted to VERL format for reward modeling
Verified ground truth answers
Ready for reinforcement learning training
🔗 Source Dataset
Original Repository
Repository: Skywork/Skywork-OR1-RL-Data
License: Apache 2.0
Paper:… See the full description on the dataset page: https://huggingface.co/datasets/sungyub/skywork-or1-math-verl.docqa-rl-verl
DocQA-RL-1.6K (VERL Format)
This dataset contains 1,591 challenging long-context document QA problems from DocQA-RL-1.6K, converted to VERL (Volcano Engine Reinforcement Learning) format for reinforcement learning training workflows.
Source: Tongyi-Zhiwen/DocQA-RL-1.6K
License: Apache 2.0
Note: This dataset maintains the original high-quality structure with user-only messages. The extra_info field has been standardized to contain only the index field for consistency with other VERL… See the full description on the dataset page: https://huggingface.co/datasets/sungyub/docqa-rl-verl.eurus-2-code-verl
Eurus-2-Code-RL (VERL Format)
This dataset contains 25,276 competitive programming problems from the Eurus-2-RL-Data dataset, filtered and converted to VERL format for reinforcement learning training workflows.
Source: PRIME-RL/Eurus-2-RL-Data
License: MIT
Note (Updated 2025-10-27): System prompts have been removed from all examples for better compatibility with other code datasets. The dataset now contains only user messages with the coding problems. See changelog for details.… See the full description on the dataset page: https://huggingface.co/datasets/sungyub/eurus-2-code-verl.code-contests-plus-verl
Code Contests Plus (VERL Format)
This dataset contains 8,432 competitive programming problems from the Code-Contests-Plus dataset, converted to VERL format for reinforcement learning applications. Each problem includes test cases validated through sandbox execution.
Source: ByteDance-Seed/Code-Contests-Plus (1x config)
License: MIT
Dataset Structure
The dataset follows the VERL format with the following fields:
data_source (string): Dataset source identifier… See the full description on the dataset page: https://huggingface.co/datasets/sungyub/code-contests-plus-verl.codev-r1-verl
CodeV-R1 VERL
📊 Dataset Summary
This dataset contains 3,133 Verilog HDL code generation problems in VERL format with standardized instruction formatting, processed from zhuyaoyu/CodeV-R1-dataset.
Key Features:
3,133 high-quality Verilog HDL problems (3,033 train + 100 validation)
Standardized instruction format with 5-section structure
JSON ground truth for equivalence checking via Sandbox Fusion
Ready for reinforcement learning training
🔗 Source… See the full description on the dataset page: https://huggingface.co/datasets/sungyub/codev-r1-verl.