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.DeepScaleR-Easy-Medium-Hard-Gemma-26B-PT-10k
DeepScaleR Easy/Medium/Hard — Gemma 4 26B-A4B PT
This dataset contains 9,900 unique, deduplicated DeepScaleR math questions for
reinforcement-learning experiments. Difficulty is defined by how often the
pretrained google/gemma-4-26B-A4B teacher solved each question across eight
temperature-1 samples under the same rule-based grader used by the RL training
pipeline.
The Hub dataset has three configurations—easy, medium, and hard—and each
configuration has a train split with 3,000… See the full description on the dataset page: https://huggingface.co/datasets/JWei05/DeepScaleR-Easy-Medium-Hard-Gemma-26B-PT-10k.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.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.deepscaler-easy10k-skyrl
deepscaler-easy10k-skyrl
Data for the short-horizon GRPO-vs-PPO study (Qwen2.5-3B, SkyRL). Parquets are in SkyRL's data.train_data / data.val_data format.
train/train.parquet — 9,800 DeepScaleR prompts, difficulty-balanced over buckets 1/5–4/5 (2,450 each); train/val.parquet (200 held-out) and train/val_x16.parquet are the original in-run val sets (legacy, do not use for reporting). train/manifest.json records the split.
eval/deepscaler_dedup500_x4.parquet — 500 deduplicated… See the full description on the dataset page: https://huggingface.co/datasets/JWei05/deepscaler-easy10k-skyrl.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-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.deepscaler-teacher-sft-vllm-official-40k
DeepScaleR teacher SFT vLLM official 40k
Generated run: exp_003_vllm_official_brainlab_2gpu.
Summary
{
"num_examples": 40300,
"sft_dir": "data/processed/deepscaler/teacher_sft/exp_003_vllm_official_brainlab_2gpu",
"parse_rate": 0.9999751861042183,
"correct_rate": 0.5728039702233251,
"format_rate": 0.005955334987593052,
"mean_reward": 0.42432258064534184,
"deepscaler_mean_reward": 0.6266997518610422,
"deepscaler_match_mean_reward":… See the full description on the dataset page: https://huggingface.co/datasets/ThunderstormXXL/deepscaler-teacher-sft-vllm-official-40k.qwen3-8b-base-deepscaler-rollouts
Qwen3-8B-Base rollouts on DeepScaleR, with verifier labels
Frozen on-policy rollouts collected for a two-branch generative-critic study. One
actor, sampled once; every downstream experiment reuses this exact batch.
Generation
actor
Qwen/Qwen3-8B-Base (chat template, enable_thinking=False)
prompts
agentica-org/DeepScaleR-Preview-Dataset, 10,000 sampled (seed 0)
samples per prompt
8
temperature / top-p
0.8 / 0.95
max new tokens
6,144 (context 8… See the full description on the dataset page: https://huggingface.co/datasets/YYYYYYibo/qwen3-8b-base-deepscaler-rollouts.deepscaler-curated
DeepScaleR Curated (8,000 problems)
A quality-screened, 8,000-problem subset of the DeepScaleR math reasoning corpus, prepared as a
drop-in training set for a Verl-based GRPO run reproducing the DeepSeek-R1 "aha moment"
experiment on a small (1.5B) pretrained model.
Source
This dataset was filtered from agentica-org/DeepScaleR-Preview-Dataset
(repo id: agentica-org/DeepScaleR-Preview-Dataset), the training corpus open-sourced by the
DeepScaleR team (AIME 1984-2023… See the full description on the dataset page: https://huggingface.co/datasets/dusersad12/deepscaler-curated.deepscaler
deepscaler
DeepScaler math problems - challenging mathematical reasoning problems.
Dataset Structure
This dataset is in Hugging Face datasets format. Load it with:
from datasets import load_dataset
dataset = load_dataset("Tyrion279/deepscaler")
DeepScaleR-Preview-Curated
DeepScaleR-Preview-Curated
DeepScaleR-Preview-Curated is a curated revision of the
agentica-org/DeepScaleR-Preview-Dataset
snapshot used for our Verl (GRPO) math-RL runs.
The published snapshot (226 entries, 220 unique problems) was
reconciled against the maintainer's revision sheet for the next release: retracted problems were dropped, duplicate
uploads were collapsed onto their first occurrence, corrected answers were taken as the authoritative ground truth, and
the… See the full description on the dataset page: https://huggingface.co/datasets/dusersad12/DeepScaleR-Preview-Curated.deepscaler-teacher-sft-vllm-official-40k-clean-v2
DeepScaleR Teacher40k Clean v2
Filtered version of ThunderstormXXL/deepscaler-teacher-sft-vllm-official-40k.
Filtering
minimum official reward: 1.0
maximum text tokens: 8192
maximum response chars: 65000
near-duplicate SimHash hamming threshold: 4
required <think>...</think> and final boxed answer after reasoning
exact text/problem/response dedupe and near problem dedupe
Counts
raw examples: 40300
kept examples: 21727
train examples: 21292
val… See the full description on the dataset page: https://huggingface.co/datasets/ThunderstormXXL/deepscaler-teacher-sft-vllm-official-40k-clean-v2.deepscaler-teacher-sft-vllm-official-40k-clean-v3-no-reward-filter
deepscaler-teacher-sft-vllm-official-40k-clean-v3-no-reward-filter
Filtered version of ThunderstormXXL/deepscaler-teacher-sft-vllm-official-40k.
Filtering
reward filter enabled: False
minimum official reward: 1.0
scoring errors rejected: False
maximum text tokens: 8192
maximum response chars: 65000
near-duplicate SimHash hamming threshold: 4
required <think>...</think> and final boxed answer after reasoning
exact text/problem/response dedupe and near problem… See the full description on the dataset page: https://huggingface.co/datasets/ThunderstormXXL/deepscaler-teacher-sft-vllm-official-40k-clean-v3-no-reward-filter.deepscaler-teacher-sft-vllm-official-40k-clean-v4-conceptual
deepscaler-teacher-sft-vllm-official-40k-clean-v4-conceptual
Filtered version of ThunderstormXXL/deepscaler-teacher-sft-vllm-official-40k.
Filtering
reward filter enabled: False
minimum official reward: 1.0
scoring errors rejected: False
maximum text tokens: 32768
maximum response chars: 200000
near-duplicate SimHash hamming threshold: 4
required <think>...</think> and final boxed answer after reasoning
exact text/problem/response dedupe and near problem dedupe… See the full description on the dataset page: https://huggingface.co/datasets/ThunderstormXXL/deepscaler-teacher-sft-vllm-official-40k-clean-v4-conceptual.DeepScaleR-Preview-Dataset-Japanese
DeepScaleR Preview Dataset (Japanese Translation)
agentica-org/DeepScaleR-Preview-Dataset の日本語翻訳版です。
データセットの説明
数学の問題、解答、解法を含むデータセットです。元のデータセットを日本語に翻訳しました。
使用方法
from datasets import load_dataset
ds = load_dataset("kfsky/DeepScaleR-Preview-Dataset-Japanese")
print(ds["train"][0])
カラム
カラム名
説明
problem
数学の問題文（日本語）
answer
解答
solution
解法の詳細（日本語）
unit
数学の単元名
ライセンス
元データセットと同じ MIT License に従います。
謝辞
元データセット:… See the full description on the dataset page: https://huggingface.co/datasets/kfsky/DeepScaleR-Preview-Dataset-Japanese.DeepScaleR-Qwen3-1.7B-2k-diverse-agreed-coded
DeepScaleR-Qwen3-1.7B-2k diverse-agreed, strategy-coded
1635 competition-math problems (the claude_agrees_gold == True subset of a
2k diverse-classified DeepScaleR pool). Each row carries Claude's worked
claude_solution plus three leak-free re-expressions of the strategy it
deploys, drawn from a shared 116-code strategy codebook.
Columns
idx — row index into agentica-org/DeepScaleR-Preview-Dataset (resume/join key).
problem, answer — the problem and gold answer.… See the full description on the dataset page: https://huggingface.co/datasets/zjhhhh/DeepScaleR-Qwen3-1.7B-2k-diverse-agreed-coded.deepscaler-1p5b-teacher-rollouts
DeepScaleR-1.5B teacher rollouts (frozen sampling)
Fixed distillation snapshot produced by sampling the frozen published RL checkpoint
agentica-org/DeepScaleR-1.5B-Preview (revision e3f524ce…) on its own published RL data
agentica-org/DeepScaleR-Preview-Dataset (revision b6ae8c60…).
The RL model was never updated — this is ordinary inference.
Construction
questions
4,096, selected deterministically by sha256(normalized_problem) lexicographic order… See the full description on the dataset page: https://huggingface.co/datasets/namezz/deepscaler-1p5b-teacher-rollouts.DeepScaleR-10k-easy
DeepScaleR-10k-easy
A 10,000-question, difficulty-balanced subset of
agentica-org/DeepScaleR-Preview-Dataset
with a non-empty worked solution for every problem. Splits: train 9,800 / val 200.
Columns
column
description
problem
the problem statement (plain text, no chat template or answer-format suffix)
answer
the grading ground truth (the form checked by a math_verify-based symbolic grader)
solution
a worked step-by-step solution (non-empty for all… See the full description on the dataset page: https://huggingface.co/datasets/JWei05/DeepScaleR-10k-easy.deepscaler-preview-kodeepscaler-1p5b-teacher-rollouts-broad
DeepScaleR-1.5B teacher rollouts — breadth variant (all trajectories kept)
Sampling of the frozen published RL checkpoint
agentica-org/DeepScaleR-1.5B-Preview (revision e3f524ce…) on its own published RL data
agentica-org/DeepScaleR-Preview-Dataset (revision b6ae8c60…). The RL model is never
updated — this is ordinary inference.
This is the breadth counterpart to
namezz/deepscaler-1p5b-teacher-rollouts.
That one sampled 4,096 questions 32 times each and kept up to 8 correct… See the full description on the dataset page: https://huggingface.co/datasets/namezz/deepscaler-1p5b-teacher-rollouts-broad.deepscale_r_math_problemsMix of agentica-org/DeepScaleR-Preview-Dataset and inclusionAI/Ring-lite-rl-data only MATH data, deduplicated
Results in total of 80k samples
DeepScaleR-Qwen3-1.7B-2k-strategy-error-200
DeepScaleR Qwen3 1.7B 2K strategy errors
This dataset contains 200 distinct questions selected from
zjhhhh/DeepScaleR-Qwen3-1.7B-2k-agreed-regraded-le5-coded
at revision 8b6e0f481bced00132c95fb631745d4992fa19fd.
Each row has one manually selected model response whose main failure is a
strategy error relative to the source row's code_hint: the response does not
materially use the hint's core route, substitutes another strategy, or omits a
decisive hinted stage in favor of an… See the full description on the dataset page: https://huggingface.co/datasets/zjhhhh/DeepScaleR-Qwen3-1.7B-2k-strategy-error-200.DeepScaleR-10k_8sampleDeepSaleR数据集中选取1k子集，采样8次
使用模型：qwen2.5-7b-insturct
使用prompt：
ITER_GEN_MULTI_TURN_STEP_PROMPTS = {
"turn0" : "{{ problem }} Let's think step by step and output the final answer within \\boxed{}.",
"turn1" : """{% if correctness %}
3. Since your initial response is self-evaluated as correct, confirm it and provide no further modifications. Put your final answer within \\boxed{}.
{% else %}
3. Since your initial response is self-evaluated as incorrect, there might be an error in the… See the full description on the dataset page: https://huggingface.co/datasets/tttonyyy/DeepScaleR-10k_8sample.deepscaler_cot