datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
UltraChat-300K-SLAM-Omni
UltraChat-300K
This dataset is prepared for the reproduction of SLAM-Omni.
This is a multi-round English spoken dialogue training dataset. For code and usage examples, please refer to the related GitHub repository: X-LANCE/SLAM-LLM (examples/s2s)
🔧 Modifications
Data Filtering: We removed samples with excessively long data.
Speech Response Tokens: We used CosyVoice to synthesize corresponding semantic speech tokens for the speech response. These tokens, represented as… See the full description on the dataset page: https://huggingface.co/datasets/worstchan/UltraChat-300K-SLAM-Omni.PersonaMem-v3
PersonaMem-v3: Toward Omni-Platform Personal Intelligence for Holistic User Understanding, Recommendation, and Agentic Tasks
Bowen Jiang, Yuan Yuan, Zhuoqun Hao, Yuchen Liu, Maohao Shen, Sihao Chen, Gregory Wornell,
Chris Callison-Burch, Lyle Ungar, Dan Roth, Qi Guo, Xiangjun Fan, Camillo J. Taylor, Hanchao Yu
A collaboration between:
Meta Recommendation Systems
University of Pennsylvania
MIT
Third release in the PersonaMem series:
PersonaMem-v1: [COLM… See the full description on the dataset page: https://huggingface.co/datasets/bowen-upenn/PersonaMem-v3.LiveSports-3K
LiveSports-3K Benchmark
News
[2025.05.12] We released the ASR transcripts for the CC track. See LiveSports-3K-CC.json for details.
Overview
LiveSports‑3K is a comprehensive benchmark for evaluating streaming video understanding capabilities of large language
and multimodal models. It consists of two evaluation tracks:
Closed Captions (CC) Track: Measures models’ ability to generate real‑time commentary aligned with the
ground‑truth ASR transcripts.
Question… See the full description on the dataset page: https://huggingface.co/datasets/stdKonjac/LiveSports-3K.BTF-3
Bench to the Future 3 (BTF-3)
1,783 pastcasting questions — 1,471 binary ("yes/no") and 312 numeric
(value-estimation) — with a state-of-the-art forecast, ground-truth resolution,
and a human-verifiable resolution explanation for every question. Designed for
reproducible evaluation of forecasting agents without hindsight bias or
web-data leakage.
BTF-3 questions were anchored to a present date in late April–late May 2026
and resolved between mid-May and early July 2026. It… See the full description on the dataset page: https://huggingface.co/datasets/BTF-2/BTF-3.ruler-300-seed42
Frozen RULER 300, seed 42
This dataset freezes the exact RULER inputs used by the
short-long-pretraining native evaluation suite.
Repository: bicycleman15/ruler-300-seed42
Rows: 6,300
Tasks: s-niah-1, s-niah-2, s-niah-3, mk1, mk2, mv, mq
Context lengths: 1024, 2048, 4096
Samples per task/length: 300
Seed: 42
Dataset SHA-256: 4d82df6f9b1f2d9c45c0a0bda8c734032e62f517b746c6351bf9c2f38335ab3d
Tokenizer SHA-256: 1f186971e25f7bda3dd6f93a100bb8fa2a6801cf8dc3807c8a8c4e45f296ab90… See the full description on the dataset page: https://huggingface.co/datasets/bicycleman15/ruler-300-seed42.qwen3.8-max-distillation-50k
Qwen3.8-Max Distillation 50K
A curated dataset of 49,772 teacher-generated traces from qwen3.8-max-preview, prepared for supervised fine-tuning and off-policy knowledge distillation.
The teacher responses are preserved as returned by the API. Where the model emitted visible <think>...</think> blocks, those blocks remain in the assistant message. Some simpler prompts received direct answers without a thinking block.
[!CAUTION]
Terms and provenance notice — not cleared for… See the full description on the dataset page: https://huggingface.co/datasets/r0b0tlab/qwen3.8-max-distillation-50k.irs-990-parsed
IRS 990 Parsed Nonprofit Database
Public relational extract of IRS Form 990 / 990-EZ / 990-PF filings, plus the colocated public files we join for address research: CMS NPPES + T-MSIS Medicare spend, FMCSA DOT carriers, OFAC SDN, FEC committees, and the IRS EO BMF.
Generated: 2026-08-17Tables: 34Rows (sum): 459,069,505License: CC0 / public domain — derived from U.S. government recordsHub: https://huggingface.co/datasets/piercewetter3/irs-990-parsed
Layout
Tables… See the full description on the dataset page: https://huggingface.co/datasets/piercewetter3/irs-990-parsed.knowchat-multi-turn-dialogues
KnowChat: Multi-Turn Human-LLM Dialogues on Knowledge Tasks
KnowChat is a dataset of 705 multi-turn human-LLM conversations collected to validate the KnowSim user simulation framework. It pairs each conversation with pre/post knowledge assessments, self-reported survey ratings, and participant background information, enabling research on information calibration -- how well LLM assistants tailor responses to users with different knowledge levels.
Dataset Summary… See the full description on the dataset page: https://huggingface.co/datasets/yjlee36/knowchat-multi-turn-dialogues.turkce-sft-qa-3.7m
🇹🇷 Turkish SFT/QA — Birleştirilmiş ve Tekrarsız Veri Seti
3,723,264 örnek. 24 açık Türkçe SFT/QA veri setinin, satır düzeyinde
tekrar temizliği ve kalite kontrolünden geçirilmiş birleşimi. Her satır hangi veri
setinden geldiğini taşır.
English: A merged, row-level deduplicated and quality-filtered collection of
24 open Turkish SFT/QA datasets (3,723,264 examples). Every row carries
its source dataset, source URL and original license.
🙏 Teşekkür /… See the full description on the dataset page: https://huggingface.co/datasets/MercanAI/turkce-sft-qa-3.7m.AIME_2000_2026_Kimi_K3
AIME 2000–2026 — Kimi K3 reasoning traces
🔄 Changelog
2026-08-08 — full re-generation. All reasoning traces were regenerated from scratch and re-verified against the official answer key.
New schema — added gen_attempts_low, gen_attempts_high; renamed gen_parsed_answer → gen_answer_int and answer_note → problem_note; removed gen_effort, gen_pass1.
New generation — only use the bare problem (v1 appended an "ANSWER:" format instruction), so traces are cleaner.… See the full description on the dataset page: https://huggingface.co/datasets/bevangelista/AIME_2000_2026_Kimi_K3.R-3-Bench R3-Bench: LLMs Struggle with Resource-Rational Reasoning under Shared Budgets
Overview
R3-Bench evaluates resource-rational reasoning when multiple problems share a
limited resource budget. This release contains the frozen benchmark data used
by the paper across three domains:
Config
Problems
Suites
Difficulty counts
math
300
50
150 easy / 100 medium / 50 hard
coding
300
50
150 easy / 100 medium / 50 hard
abstract_reasoning
300
50
150 easy /… See the full description on the dataset page: https://huggingface.co/datasets/R-3-Bench/R-3-Bench.minimax-m3-deepsearchqa-skill-eval
MiniMax M3 DeepSearchQA Skill Eval
Evaluates minimax/minimax-m3 on google/deepsearchqa using a Pi agent, You.com MCP tools, and a research skill optimized for this harness, model, and tool surface.
MiniMax M3 Medium Reasoning with the You.com research skill reached 74.85% adjusted F1 on DeepSearchQA, above the paper's GPT-5 High Reasoning F1 result. Public artifacts are available for inspection and reproduction.
Links
GitHub:… See the full description on the dataset page: https://huggingface.co/datasets/youdotcom/minimax-m3-deepsearchqa-skill-eval.UltraChat-300K-SLAM-Omni
UltraChat-300K
This dataset is prepared for the reproduction of SLAM-Omni.
This is a multi-round English spoken dialogue training dataset. For code and usage examples, please refer to the related GitHub repository: X-LANCE/SLAM-LLM (examples/s2s)
🔧 Modifications
Data Filtering: We removed samples with excessively long data.
Speech Response Tokens: We used CosyVoice to synthesize corresponding semantic speech tokens for the speech response. These tokens, represented as… See the full description on the dataset page: https://huggingface.co/datasets/mwei/UltraChat-300K-SLAM-Omni.Qwen3-Math-Eval
Qwen3 Math Evaluation Suite
Greedy (temperature 0) outputs from Qwen3 1.7B / 4B / 8B / 14B on nine math-reasoning benchmarks across output-token budgets {2k, 4k, 8k, 16k, 32k}. 1,417,388 predictions over 186 model-by-dataset-by-budget cells, each with the full reasoning trace, the extracted answer, and strict and answer-forced correctness labels.
On standard MATH (MATH-500, Hendrycks MATH test, competition MATH) these models are saturated at 16k: the 4B is at or above 0.94 and… See the full description on the dataset page: https://huggingface.co/datasets/ssubhnil/Qwen3-Math-Eval.FlowBench
FlowBench
Dataset ID: jwu323/FlowBench
FlowBench is a tool-use benchmark for deterministic business operations
workflows. Each task asks an agent to compose Python tools over synthetic
customers, products, orders, returns, inventory, support tickets, FX rates, and
SLA policies.
Scope note: this dataset is a business-operations tool-composition benchmark. It
is unrelated to prior workflow-guided planning or workflow-generation benchmarks
that also use the FlowBench name.
This… See the full description on the dataset page: https://huggingface.co/datasets/jwu323/FlowBench.2d_3d_seq_path_spatial_reasoning
Spatial Reasoning Dataset
A synthetic dataset of Hamiltonian path puzzles with rich chain-of-thought reasoning, designed for training and evaluating spatial reasoning in language models.
Overview
Each sample presents a grid-based puzzle where the solver must find a path visiting every cell exactly once, moving only up/down/left/right (plus above/below for 3D). Puzzles span 2D grids (3x3 to 8x8) and 3D cubes (3x3x3 to 4x4x4), covering solvable, impossible, and multi-turn… See the full description on the dataset page: https://huggingface.co/datasets/eousphoros/2d_3d_seq_path_spatial_reasoning.pubmedqa-recursive-llm-degradation-qwen2.5-3b
PubMedQA Recursive LLM Degradation — Qwen2.5-3B
This repository contains synthetic biomedical question-answering data
and model predictions generated as part of a study of recursive
fine-tuning and model degradation.
Base Model
Qwen/Qwen2.5-3B
Source Dataset
The experiments use the PubMedQA dataset:
qiaoxin/PubMedQA
This repository contains generated/derived research artifacts and does
not redistribute the original PubMedQA dataset in its entirety.… See the full description on the dataset page: https://huggingface.co/datasets/chrislimbe/pubmedqa-recursive-llm-degradation-qwen2.5-3b.Maux-Persian-SFT-30k
Maux-Persian-SFT-30k
Dataset Description
This dataset contains 30,000 high-quality Persian (Farsi) conversations for supervised fine-tuning (SFT) of conversational AI models. The dataset combines multiple sources to provide diverse, natural Persian conversations covering various topics and interaction patterns.
Dataset Structure
Each entry contains:
messages: List of conversation messages with role (user/assistant/system) and content
source: Source dataset… See the full description on the dataset page: https://huggingface.co/datasets/xmanii/Maux-Persian-SFT-30k.cross-lingual-pitfalls
Cross-Lingual Pitfalls
Cross-Lingual Pitfalls is a fixed, failure-focused dataset from the ACL 2025 paper "Cross-Lingual Pitfalls: Automatic Probing Cross-Lingual Weakness of Multilingual Large Language Models." It contains 6,713 bilingual English-to-target-language question pairs across 16 target languages. The paper's search-based multilingual LLM evaluation method uses beam search and LLM-based simulation to discover cases where a model answers correctly in English but fails… See the full description on the dataset page: https://huggingface.co/datasets/xzx34/cross-lingual-pitfalls.acc_rd_s1-gpqa
Dataset Card for GPQA
GPQA is a multiple-choice, Q&A dataset of very hard questions written and validated by experts in biology, physics, and chemistry. When attempting questions out of their own domain (e.g., a physicist answers a chemistry question), these experts get only 34% accuracy, despite spending >30m with full access to Google.
We request that you do not reveal examples from this dataset in plain text or images online, to reduce the risk of leakage into foundation model… See the full description on the dataset page: https://huggingface.co/datasets/stewy33/acc_rd_s1-gpqa.cs_squad-3.0
Dataset Card for Czech Simple Question Answering Dataset 3.0
This a processed and filtered adaptation of an existing dataset. For raw and larger dataset, see Dataset Source section.
Dataset Description
The data contains questions and answers based on Czech wikipeadia articles.
Each question has an answer (or more) and a selected part of the context as the evidence.
A majority of the answers are extractive - i.e. they are present in the context in the exact form. The… See the full description on the dataset page: https://huggingface.co/datasets/fewshot-goes-multilingual/cs_squad-3.0.longctx30
longctx30
Thirty long-context prompts for benchmarking LLM inference throughput. Each
prompt is about 10,000 input tokens and asks for about 1,500 output tokens,
which is long enough that decode time dominates and tokens per second is a
meaningful number.
Built for the article Learning inference: How to host and improve the token
speed of an LLM, where it is the benchmark set for Gemma 4 31B on a single
B300. The data and code the article uses are in this repository:
file… See the full description on the dataset page: https://huggingface.co/datasets/abhijithneilabraham/longctx30.Chinese-Qwen3-235B-Thinking-2507-Distill-100k
📌 Note: The English translation of this dataset card is provided below.
Chinese-Qwen3-235B-Thinking-2507-Distill-100k
Dataset Summary
Chinese-Qwen3-235B-Thinking-2507-Distill-100k 是一个包含约 100k 条高质量中文推理与指令数据的数据集，由 Qwen-3-235B-A22B-Thinking-2507（官方 Thinking 模式，上下文长度 32K）蒸馏生成。
该数据集覆盖了多个重要领域：
数学与工程任务（Mathematics, Applied Math, Advanced Math）
通用知识与写作（General Knowledge, Language & Writing）
技术与编程（Technology & Programming）
商业与经济（Business & Economics）… See the full description on the dataset page: https://huggingface.co/datasets/Jackrong/Chinese-Qwen3-235B-Thinking-2507-Distill-100k.Llama_3.1-8B-Instruct-Self-CalibrationThe official repository which contains the code and pre-trained models/datasets for our paper Efficient Test-Time Scaling via Self-Calibration.
🔥 Updates
[2025-3-3]: We released our paper.
[2025-2-25]: We released our codes, models and datasets.
🏴 Overview
We propose an efficient test-time scaling method by using model confidence for dynamically sampling adjustment, since confidence can be seen as an intrinsic measure that directly reflects model… See the full description on the dataset page: https://huggingface.co/datasets/HINT-lab/Llama_3.1-8B-Instruct-Self-Calibration.SpaceOmicsBench-v3
SpaceOmicsBench v3
A Multi-Omics AI Benchmark for Spaceflight Biomedical Data
SpaceOmicsBench v3 provides standardized ML and LLM evaluation infrastructure for spaceflight biomedical data from 4 human spaceflight missions (NASA Twins Study, Inspiration4, JAXA cfRNA, Axiom-2).
Dataset Structure
ML Track (Track A)
tasks/track_a/ — Task definitions (J1: phase classification, J2: clock acceleration)
tasks/track_c/ — Feature-level task definitions (C1:… See the full description on the dataset page: https://huggingface.co/datasets/jang1563/SpaceOmicsBench-v3.k3-sft-cc0-flan
Dataset Card for K3 SFT CC0 FLAN
844-row Kimi K3 synthetic instruction-tuning shard built from DPI-traced CC0/public-domain
FLAN prompts in the Tülu mix. Four overlapping Hub configs expose different cohort
views; adaptive is the recommended default for quality-conscious SFT mixing.
Dataset Details
Curated by: Training Datasmith
Teacher: kimi-k3 via deltafin (local inference)
Languages: English prompts; translation pairs include German, Spanish, Czech, Igbo… See the full description on the dataset page: https://huggingface.co/datasets/Training-Datasmith/k3-sft-cc0-flan.AIME_1983_2026_Kimi_K3
AIME 1983–2026 — Kimi K3 reasoning traces
🔄 Changelog
2026-08-08 — full re-generation. All reasoning traces were regenerated from scratch and re-verified against the official answer key.
New schema — added gen_attempts_low, gen_attempts_high; renamed gen_parsed_answer → gen_answer_int and answer_note → problem_note; removed gen_effort, gen_pass1.
New generation — only use the bare problem (v1 appended an "ANSWER:" format instruction), so traces are cleaner.… See the full description on the dataset page: https://huggingface.co/datasets/bevangelista/AIME_1983_2026_Kimi_K3.lean-proof-or-refute-300
Lean Proof-or-Refute 300
Lean Proof-or-Refute 300 is a compact collection of 300 formal reasoning
problems grounded in Lean 4 and Mathlib. Each problem starts from a verified
Mathlib theorem, makes one small numerical or operator mutation, and asks the
model to return either:
a Lean certificate proving the mutated proposition; or
a Lean certificate proving the exact negation of the complete proposition.
The model receives the related source theorem, a bounded source excerpt… See the full description on the dataset page: https://huggingface.co/datasets/xlr8harder/lean-proof-or-refute-300.threejs-gamecode-instruct-v3-ultra
Three.js GameCode Instruct v3 Ultra
This is a large synthetic/original instruction dataset for training or testing LLM behavior around Three.js, browser game development, gameplay programming, debugging, optimization, architecture, and general coding.
Important note
This dataset is synthetic and programmatically generated from original templates. It is designed as a useful starting point for experiments, not as a fully hand-curated gold-standard benchmark.
No… See the full description on the dataset page: https://huggingface.co/datasets/agagasf123123/threejs-gamecode-instruct-v3-ultra.biosum-cuh
BioSum-CUH
A Biography Summarization Benchmark with Token-Level Correctness, Uncertainty, and Hallucination Annotations
Paper: UT-ACA: Uncertainty-Triggered Adaptive Context Allocation for Long-Context Inference | Code: github.com/Tommy307/UT-ACA
BioSum-CUH is a benchmark for studying factual generation over biography
contexts. It combines biography-based question answering and structured
summarization with token-aligned model predictions, final-layer attention
activations… See the full description on the dataset page: https://huggingface.co/datasets/Saria307/biosum-cuh.