datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
Polaris
Polaris Dataset
🌟 [CVPR24] Polos: Multimodal Metric Learning from Human Feedback for Image Captioning
Accepted at CVPR 2024
🌐 project page
📄 arXiv
🤗 Dataset
Establishing an automatic evaluation metric that closely aligns with human judgements is essential for the effective development of image captioning models. Data-driven metrics have recently gained prominence in this field, demonstrating a stronger correlation with human judgements than classic metrics such as CIDEr and… See the full description on the dataset page: https://huggingface.co/datasets/yuwd/Polaris.Polaris-Dataset-53K
Overview
Training dataset for Polaris Preview models. The dataset is filtered from DeepScaleR-Preview-Dataset and AReal-boba-Data
Format
Each row in the jsonl file contains:
problem: The input problem.
answer: The answer to the problem
difficulty: The pass rate of the problem estimated by Deepseek-R1-distill-Qwen-7B
Citation
@misc{Polaris2025,
title = {POLARIS: A Post-Training Recipe for Scaling Reinforcement Learning on Advanced Reasoning Models}… See the full description on the dataset page: https://huggingface.co/datasets/POLARIS-Project/Polaris-Dataset-53K.polaris-bench
The Cartesian Shortcut: Re-evaluate Vision Reasoning in Polar Coordinate Space
Polaris-Bench: Official Evaluation Dataset
Xia Hu1,
Zhenrui Yue1,
Brian Potetz1,
Howard Zhou1,
Leonidas Guibas1,2,
Chun-Ta Lu3,
Zhicheng Wang1
1Google DeepMind 2Stanford University 3Google Research
Overview
As current Multimodal Large Language Models (MLLMs) rapidly saturate canonical visual reasoning benchmarks, a key… See the full description on the dataset page: https://huggingface.co/datasets/google/polaris-bench.maxrl_qwen3_4B_base_polaris_rollouts
MaxRL Qwen3-4B-Base training rollouts (POLARIS math prompts)
Every training rollout from an online RL run, with exact token ids, sampling
log-probs, and raw rewards — usable as a replay buffer to study off-policy RL
for LLM reasoning completely offline.
The run: Qwen3-4B-Base trained with the maxRL advantage estimator
(A = (r - mean)/(mean + eps), group mean over 16 rollouts per prompt;
maxRL paper) and a pure REINFORCE loss
(L = -A * log pi; no importance ratio, no clipping, no… See the full description on the dataset page: https://huggingface.co/datasets/ftajwar/maxrl_qwen3_4B_base_polaris_rollouts.qwen3-1.7b-polaris-fp8-rollouts-20260912
Qwen3-1.7B Polaris FP8 rollouts
Snapshot of three FP8 rollout datasets taken on 2026-09-12. The model is Qwen3-1.7B-Base. Original Parquet files and attempt, file, and checkpoint-lineage metadata are preserved without rewriting.
Configuration
Training batches
Training responses
Validation responses
Total bytes
maxrl_strict
101
827392
144320
6654288117
maxrl_permissive
107
876544
144320
8168064932
dppo
126
1032192
173184
6733682534
Provenance and… See the full description on the dataset page: https://huggingface.co/datasets/steviel/qwen3-1.7b-polaris-fp8-rollouts-20260912.Polaris-Qwen3-4B-Instruct-eval-32k-64_partialpolaris_imagereward_v3rloo_qwen3_1p7B_base_polaris_rolloutspolaris_imagerewardpolaris-ecir-v1
ECIR v1
ECIR is one of six datasets in Polaris: Learning to Generate Table Descriptions from Retrieval
Feedback, alongside aw, arctic, lter, wikitables, and
wtr.
It holds 2,100 tables published on the US government open data portal and 12 keyword queries over
them. For each query–table pair, a person decided how well that table answers that query and gave it
a score; those scores are the relevance judgments, and they live in qrels.csv. Given a query, a
system ranks the 2,100… See the full description on the dataset page: https://huggingface.co/datasets/anhaidgroup/polaris-ecir-v1.polaris-wikitables-v2
WikiTables v2
WikiTables is one of six datasets in Polaris: Learning to Generate Table Descriptions from
Retrieval Feedback, alongside aw, arctic, lter, ecir,
and wtr.
It holds 3,361 tables scraped from Wikipedia articles and 57 keyword queries over them. For each
query–table pair, a person scored how well that table answers that query; those scores are the
relevance judgments, and they live in qrels.csv.
The tables have no names. What describes a table is its column names and… See the full description on the dataset page: https://huggingface.co/datasets/anhaidgroup/polaris-wikitables-v2.polaris_imagereward_v2leaderboard-datamonitorability-as-a-free-gift-data-training-data
Monitorability as a free gift training data reordered, or normalized during packaging.
Configurations
Config
Rows
Purpose
Original file
all
18,591
Main all-domain experiment
combined_dataset.parquet
no_if
13,591
All-domain experiment without instruction following
combined_dataset_noif.parquet
instruction_following
5,000
Instruction-following experiments
instruction_following_ai2_5000.parquet
math
5,000
Main math experiments
skywork_math.parquet… See the full description on the dataset page: https://huggingface.co/datasets/polaris-73/monitorability-as-a-free-gift-data-training-data.Polaris-Hard
Polaris-Hard
A random subset of Polaris-Dataset-53K, with 3,200 distinct math problems in a single train split.
Original difficulty
Source pool
Eligible pool
Selected
Share
0/8
15,368
9,331
2,000
62.5%
1/8
6,956
4,929
1,200
37.5%
Total
22,324
14,260
3,200
100%
Sampling
Source revision: 296f8e34132e63f4a1d70e0dcc8bddebb43f03e4.
Seed: 42. Uniform random sampling without replacement within each group, followed by a deterministic shuffle of the… See the full description on the dataset page: https://huggingface.co/datasets/hi-todayis-jh/Polaris-Hard.polaris-acemath-gemini-rubrics-v2Polaris-cleanfaithful-thinking-draft
Dataset Card for Thinking Draft Faithfulness Evaluation
This dataset accompanies the paper "Measuring the Faithfulness of Thinking Drafts in Large Reasoning Models".
Dataset Description
The Faithful Thinking Draft dataset is designed to evaluate how faithfully language models follow their own thinking drafts. It contains benchmarks for two key aspects of reasoning faithfulness:
Intra-Draft Faithfulness: Tests how consistently models follow their own reasoning steps when… See the full description on the dataset page: https://huggingface.co/datasets/polaris-73/faithful-thinking-draft.POLARIS
POLARIS
POLARIS is a prompt-only dataset release for long-form story generation. It contains the training prompts used for the POLARIS story-writing models together with the official test prompts used in evaluation.
This release is intentionally narrow: it is designed to support reproducibility for prompt-based evaluation and generation experiments without releasing copyrighted story text or training-time reasoning traces.
What is included
The dataset has two… See the full description on the dataset page: https://huggingface.co/datasets/rishanthrajendhran/POLARIS.polaris-lter-v1
LTER v1
LTER is one of six datasets in Polaris: Learning to Generate Table Descriptions from Retrieval
Feedback, alongside aw, arctic, ecir, wikitables, and
wtr.
It holds 2,015 tables from the top-downloaded collections of the Long-Term Ecological Research sites
in the Environmental Data Initiative (EDI) — bird surveys, forest phenology, reef colonisation,
cattle records — and 15 keyword queries over them. For each query–table pair, a person decided
whether that table answers… See the full description on the dataset page: https://huggingface.co/datasets/anhaidgroup/polaris-lter-v1.Polaris-clean-v2Polaris-1-8-3200
Polaris-1-8-3200
A random subset of Polaris-Dataset-53K, with 3,200 distinct math problems in a single train split.
Original difficulty
Source pool
Eligible pool
Selected
Share
1/8
6,956
4,929
3,200
100.0%
Total
6,956
4,929
3,200
100%
Sampling
Source revision: 296f8e34132e63f4a1d70e0dcc8bddebb43f03e4.
Seed: 42. Uniform random sampling without replacement within each group, followed by a deterministic shuffle of the combined selection.… See the full description on the dataset page: https://huggingface.co/datasets/hi-todayis-jh/Polaris-1-8-3200.pretrain_polarispolaris-53k-repaired
POLARIS-53K, label-repaired
49,289 of the 53,291 rows in
POLARIS-Project/Polaris-Dataset-53K,
with 4,580 stored answers corrected and 4,002 rows removed as unrepairable.
Measurements on the source set put its bad-label rate at roughly 15.9%
[14.3, 17.6] (two independent detectors agreeing on a 2,000-row sample).
Mislabelled rows are not uniformly distributed: they concentrate in the problems
models fail, which is exactly where a difficulty-calibration pipeline looks.… See the full description on the dataset page: https://huggingface.co/datasets/joanvelja/polaris-53k-repaired.grpo_qwen3_1p7B_base_polaris_rolloutspolaris-alpha-1000xThis is a non-reasoning dataset created using Polaris Alpha (likely an alpha version of a new ChatGPT) from OpenAI. Some of these questions are from reedmayhew and the rest were generated.
Most of the questions cover the following topics: Web Development, Logic, Math, Embedded Systems, Web Design and Python Scripting.
The dataset is meant for creating distilled versions of Polaris Alpha by fine-tuning already existing open-source LLMs.
polaris-aw-v2
AW v2
AW is one of six datasets in Polaris: Learning to Generate Table Descriptions from Retrieval
Feedback, alongside arctic, lter, ecir, wikitables, and
wtr.
It holds 96 tables from a version of AdventureWorks whose column names are cryptic — BusEntId,
STRGUID, JobTtl — and 15 keyword queries over them. For each query–table pair, a person decided
whether that table answers that query; those decisions are the relevance judgments, and they live in
qrels.csv.
Each Polaris dataset… See the full description on the dataset page: https://huggingface.co/datasets/anhaidgroup/polaris-aw-v2.polaris-ecir-v2
ECIR v2
ECIR is one of six datasets in Polaris: Learning to Generate Table Descriptions from Retrieval
Feedback, alongside aw, arctic, lter, wikitables, and
wtr.
It holds 2,100 tables published on the US government open data portal and 12 keyword queries over
them. For each query–table pair, a person decided how well that table answers that query and gave it
a score; those scores are the relevance judgments, and they live in qrels.csv. Given a query, a
system ranks the 2,100… See the full description on the dataset page: https://huggingface.co/datasets/anhaidgroup/polaris-ecir-v2.maxrl_qwen3_1p7B_base_polaris_rolloutsPOLARIS-Gov
POLARIS-Gov: Political Orientation Longitudinal Analysis Resource for International States
Dataset Statistics
Countries Tracked: 171 Sovereign Nations (160 in effective modeling universe)
Observations: 3,373 country-year rows
Temporal Coverage: 2000–2019
Treatment Variable: ideology_score (Unified ideological score of governing coalitions)
Outcome Variables: GDP growth, GDP per capita growth
Control Variables: 18 World Bank Economic Indicators
Panel Type:… See the full description on the dataset page: https://huggingface.co/datasets/Maybe-Heisenberg-07/POLARIS-Gov.