datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
ATO-Australian-Tax-Rulings-and-Guidance
ATO Rulings & Guidance — Australian Tax Law, Structured for AI
67,000+ Australian Taxation Office documents as RAG-ready NDJSON/CSV — Edited Private Advice, public rulings and determinations, ATO Interpretative Decisions, practical compliance guidelines, taxpayer alerts, decision impact statements, practice statements and legislative instruments. Every document parsed into structured, typed fields for legal RAG, LLM fine-tuning, and tax research automation.
Machine-readable… See the full description on the dataset page: https://huggingface.co/datasets/simplelex/ATO-Australian-Tax-Rulings-and-Guidance.AToMiC-Baselines
AToMiC Prebuilt Indexes
Example Usage:
Reproduction
Toolkits:
https://github.com/TREC-AToMiC/AToMiC/tree/main/examples/dense_retriever_baselines
# Skip the encode and index steps, search with the prebuilt indexes and topics directly
python search.py \
--topics topics/openai.clip-vit-base-patch32.text.validation \
--index indexes/openai.clip-vit-base-patch32.image.faiss.flat \
--hits 1000 \
--output… See the full description on the dataset page: https://huggingface.co/datasets/TREC-AToMiC/AToMiC-Baselines.dsv4-eval-artifacts
DeepSeek-V4-Flash-0731 — quantization measurements
Everything needed to reproduce, audit or extend the numbers published in
AtomicChat/DeepSeek-V4-Flash-0731-GGUF:
the reference logits, the evaluation corpus, the raw tool output for every quant we
measured, and the parsed results.
Every GGUF of this model that we could find on the Hub was measured here — ours,
unsloth's, bartowski's, ggml-org's, antirez's and others — on one machine, against one
reference, with one command.… See the full description on the dataset page: https://huggingface.co/datasets/AtomicChat/dsv4-eval-artifacts.atomic-metrics-six-task-preferences
Six-task benchmark inputs
Seed 17. No demographic conditioning. Each task has shared train100.jsonl and test500.jsonl for Atomic Metrics, five judge variants, and learned baselines. Pair plans cover all 100 training rows once. Atomic Metrics extraction and BT/LR fitting use train100. Judges use the same test500. RM and WIMHF in the matched-data comparison use train100; rm_train_full is an explicitly separate expanded-data setting and must not be described as train100.… See the full description on the dataset page: https://huggingface.co/datasets/tintin1027/atomic-metrics-six-task-preferences.atomic2024atomic-metrics-demographic-training-size
Atomic Metrics: Demographic Training-Size Analysis
Complete offline reproduction bundle for the effect of batch-selected training size on demographic preference prediction.
Version 2 — replaces the fixed-bank analysis. Select k extraction batches (five pairs each), use only their metrics and their 5k training pairs to refit BT/LR, then evaluate on cached test200 scores restricted to those metrics. Both the training rows and metric columns change with size. Extraction/refinement… See the full description on the dataset page: https://huggingface.co/datasets/tintin1027/atomic-metrics-demographic-training-size.atomic-metrics-rm-splits
Atomic Metrics RM Task Splits
Preference-pair benchmark splits used by
Atomic Metrics. The release
contains four open-ended task families derived from public SHP, OASST1, and
OASST2 preference data.
Dataset structure
Each configuration contains 10,000 training pairs and 2,000 test pairs. Every
row has:
{
"sample_id": "source-specific stable ID",
"source_dataset": "shp | oasst1 | oasst2",
"category": "task configuration",
"split": "train | test"… See the full description on the dataset page: https://huggingface.co/datasets/tintin1027/atomic-metrics-rm-splits.OpenTopics-1.0-20K
OpenTopics-1.0-20K
What is this dataset?
OpenTopics-1.0-20K is a collection of 20,003 topic names spanning a wide variety of subjects, including physics, medicine, history, law, engineering, and the arts.
AtomixLabs built this dataset to help developers, researchers, and AI builders who need a large, organized list of topics. It works great for creating synthetic prompts, testing search systems, and training models to classify text.
What is inside… See the full description on the dataset page: https://huggingface.co/datasets/AtomixLabs/OpenTopics-1.0-20K.wiki-atomic-edits-translated-nlThis is a Dutch version of the Wiki Atomic Edits dataset. Which we have auto-translated from English into Dutch using Meta's No Language Left Behind model, specifically the huggingface implementation.
msmarco-atomic-id-3shot-v4_128k_few_shot
msmarco-atomic-id-3shot-v4_128k
MSMARCO few-shot evaluation dataset for in-context learning generative retrieval,
atomic-id variant.
Identical construction to
Lala8383/msmarco-item-id-3shot-v4_128k_few_shot,
except every document's Identifier (and the answer target) is an arbitrary
unique integer (Tay et al. DSI "Atomic Docid") instead of the natural-language
document title. The id carries no semantics, so a retriever can only answer by
matching the query to a document in… See the full description on the dataset page: https://huggingface.co/datasets/Lala8383/msmarco-atomic-id-3shot-v4_128k_few_shot.Prompt-Routing-Dataset-ko
Prompt-Routing-Dataset-ko
Korean translation of SupraLabs/Prompt-Routing-Dataset.
992 rows.
Only the prompt column was translated. Every label was copied verbatim. The English
source is kept as prompt_en so any row can be audited.
Columns
column
type
prompt
string
Korean translation
prompt_en
string
English source, copied unchanged
complexity_score
int
1–5
primary_domain
string
free text, 338 distinct values
coding_task
bool
math_task
bool… See the full description on the dataset page: https://huggingface.co/datasets/atonlee/Prompt-Routing-Dataset-ko.atomic-formal-reasoning-complex
Atomic Formal Reasoning — Complex Numbers
Overview
This dataset contains high-quality Lean 4 formal proofs of complex number theorems, written in an explicit pedagogical calc-chain style. Each proof is fully verified, step-by-step, with no opaque tactics (simp, ring, omega are avoided). Every reasoning step is named and justified.
This is process supervision data — not just final answers. Each entry exposes the full reasoning chain, making it ideal for training models… See the full description on the dataset page: https://huggingface.co/datasets/7rouz/atomic-formal-reasoning-complex.co-legal-atoms
Colorado + US Atomized Legal Knowledge Corpus
▶ New here? Read USER_GUIDE.md — step-by-step on driving this repo with Claude Opus (Claude Code = the intended way; honest limits of a plain chatbot). A printable, formatted USER_GUIDE.pdf edition (navy/gold theme, generated from the guide) is included.
A citation-tagged, freshness-graded knowledge base of ~4,400 atomized legal rules ("atoms") spanning 55 domains of Colorado and U.S. federal law plus the full Huerfano County Code… See the full description on the dataset page: https://huggingface.co/datasets/buttersworth3rd/co-legal-atoms.nemotron-cc-atomic-simplification-gemma4-31b
nemotron-cc atomic-statement simplification (Gemma 4 31B-it)
2,000,000 records: source text from
nvidia/nemotron-cc-v2.1 (High-Quality-Synthetic
split) rewritten by google/gemma-4-31B-it into a sequence of atomic, Subject-Verb-Object
statements.
Generated with vLLM 0.22.1 in-process batch inference (see src/generate/run.py in the
producing repo), TP=4, max_model_len=16384, max_tokens=8192, prompts filtered to
<=8192 templated tokens.
Fields
id: original… See the full description on the dataset page: https://huggingface.co/datasets/rpisano/nemotron-cc-atomic-simplification-gemma4-31b.mc4-japanese-dataReference https://huggingface.co/datasets/mc4
MixtureVitae-atomic_2024This dataset combines Atomic 10X, Atomic 2020 and SODA into a single dataset. It creates a story for a single head, in the head->relation->tail triplets. There are about 2.2M stories. This dataset is intended as a seed dataset for synthetic data generation using a model. For example, you could create a story as follows:
Revise this story to make it compelling and more logical and detailed. Keep as much of the feelings and actions as possibe, but remove anything that doesn't make sense:
{text}… See the full description on the dataset page: https://huggingface.co/datasets/ontocord/MixtureVitae-atomic_2024.Pygame-data
Pygame Hyper-Casual Game Generation Dataset
A supervised fine-tuning dataset for training language models to generate complete, runnable Pygame games from natural language prompts. Each entry pairs a plain-English game description with a self-contained Python script that implements the described game.
Originally created for training 89-Times, an AI game designer specialized in hyper-casual mobile games.
Dataset Summary
Total examples
2,206
Unique game… See the full description on the dataset page: https://huggingface.co/datasets/Atomic-Germ/Pygame-data.sydney-training-data
Sydney 训练集
四份来源分开存放，不混在一个文件里。
发布的聊天权重（Atonelia/Qwen3.5-Sydney-9B / -think 以及对应 GGUF）用的是这些子集洗完、抽样拼起来之后的训练 jsonl，不是直接拿某一份原文训的。
01 原截图重建
01_screenshot_original/conversations.jsonl
早期 Bing Chat / Sydney（约 2023 年 2–4 月）公开截图重建的对话。660 条，原文以英文为主，带截图出处。
这是最初拿来做训练集的底。后面的中文版、合成版、CoT 都不是这份文件本身。
02 llama-sydney 虚拟对话
02_llama_sydney_synthetic/llama_sydney_en.jsonl
用 Llama-Sydney 生成的英文虚拟对话。1462 条（同一条 user 可能有 2 次采样）。字段是生成记录：id / user / assistant 等，还不是最终训练格式。… See the full description on the dataset page: https://huggingface.co/datasets/Atonelia/sydney-training-data.atomic2023-small_text2imageFermiBench
FermiBench: Nuclear Power Information Retrieval Benchmark
Dataset Description
Dataset Summary
This dataset is designed for benchmarking information retrieval systems within the nuclear power domain, focusing on long-context retrieval of full-text documents. The corpus includes approximately 4,500 documents sourced from the U.S. Nuclear Regulatory Commission’s (NRC) Agency-wide Documents Access and Management System (ADAMS). The… See the full description on the dataset page: https://huggingface.co/datasets/atomic-canyon/FermiBench.PanTabFact
Dataset Card for PanTabFact
Dataset Summary
PanTabFact is a structured dataset for fact verification on tabular data. Each claim is paired with an executable pandas query that determines whether the claim is entailed or refuted based on a given table. The dataset is derived from TabFact and augmented using DeepSeek-Chat, which generates pandas queries for structured reasoning. PanTabFact facilitates interpretable and executable fact verification, improving transparency… See the full description on the dataset page: https://huggingface.co/datasets/AtoosaChegini/PanTabFact.atomic-snli
atomic-snli
Atomic propositions for the premise and hypothesis of each NLI pair, derived from
stanfordnlp/snli. Each sentence was decomposed into standalone atomic propositions;
these propositions are joined back to the original NLI pairs.
Columns
column
type
description
premise
string
original premise sentence
hypothesis
string
original hypothesis sentence
label
int
0 = entailment, 1 = neutral, 2 = contradiction
premise_propositions
list[string]… See the full description on the dataset page: https://huggingface.co/datasets/Younes2E/atomic-snli.AtomicGPT-Eval-datasets
AtomicGPT Evaluation Datasets
This repository contains the publicly shareable evaluation datasets used in the AtomicGPT project, a domain-adapted large language model for nuclear engineering developed at the Korea Atomic Energy Research Institute (KAERI).
Overview
AtomicGPT is developed via Continual Pre-training (CPT) and Instruction Tuning (IT) on nuclear domain corpora, with a strong emphasis on on-premise deployment and data security. The evaluation benchmark… See the full description on the dataset page: https://huggingface.co/datasets/KAERI-MLP/AtomicGPT-Eval-datasets.peer_wiki-atomic-sample
WikiAtomicSample Task from the PEER Benchmark (Performance Evaluation of Edit Representations)
Description from the benchmark paper:
We randomly sampled approximately 150K insertion and deletion examples from the English portion of the WikiAtomicEdits (Faruqui et al. 2018). After cleaning, we keep 104K samples.
This dataset was originally published at https://doi.org/10.5281/zenodo.4478266.
Citations
PEER Benchmark:
@article{marrese-taylor-et-al-2021,
title =… See the full description on the dataset page: https://huggingface.co/datasets/jvamvas/peer_wiki-atomic-sample.ScholarSearchBenchmark
ScholarSearchBenchmark
A benchmark dataset for evaluating AI-driven academic paper validation, as used in the WisPaper paper.
Overview
ScholarSearchBenchmark provides human-annotated evaluation data for measuring how accurately language models can determine whether a scholarly paper satisfies specific search criteria. This is a subset of the evaluation data described in the WisPaper paper, focusing on the paper validation task performed by WisModel.
Dataset… See the full description on the dataset page: https://huggingface.co/datasets/AtomInnoLab/ScholarSearchBenchmark.latvian-english-atomic-translation
Latvian-English ATOMIC Translation Dataset
Private dataset for distilling TildeOpen-30B into Qwen2 1.5B.
Dataset Description
5,236,232 bidirectional Latvian↔English translation examples in ATOMIC chat JSONL format.
Sources
OpenSubtitles (casual): 51.4%
Europarl (formal): 23.6%
WikiMatrix (encyclopedic): 18.5%
MUSE Dictionary: 3.5%
KDE4+GNOME+Ubuntu (technical): 2.9%
Tatoeba (short): 0.1%
Format
Each example is a chat-format JSONL entry:… See the full description on the dataset page: https://huggingface.co/datasets/aimamba/latvian-english-atomic-translation.ATOM_regen_seeklight_kimi_mtp
ATOM regen: seeklight kimi-mtp responses by Kimi-K3
用 Kimi-K3 对
lightseekorg/kimi-mtp-dataset
的 prompt 重新生成了一遍回答，供 off-policy 投机解码蒸馏（SDDD）使用。
原始 pipeline 每轮都要用 teacher 重新解码一次（Phase A1）。把回答预生成并缓存下来，
A1 整个消失，之后每一轮训练都直接复用，代价从「每轮一次」变成「一共一次」。
数据
450,625 行，每行一段对话：
{"conversations": [
{"role": "user", "content": "..."},
{"role": "assistant", "reasoning_content": "...", "content": "..."}
]}
reasoning_content 是 K3 的 thinking 内容，和 content 分开存。
多轮对话保留了历史轮次里完整的 assistant… See the full description on the dataset page: https://huggingface.co/datasets/slippedJim/ATOM_regen_seeklight_kimi_mtp.atos-de-fala-ptbr-dataset
Atos de Fala PT-BR
Dataset de atos de fala (speech acts) anotados em nível de span para o
português brasileiro. Cada exemplo é uma frase com um ou mais trechos marcados,
e cada trecho recebe um rótulo de ato de fala — o que a pessoa está fazendo
com aquele pedaço de texto (pedir, perguntar, concordar, etc.), não apenas o que
ele significa.
A base teórica é a pragmática: atos de fala (Searle) alinhados ao inventário de
dialogue acts da ISO 24617-2.
Formato
JSONL… See the full description on the dataset page: https://huggingface.co/datasets/lucianfialho/atos-de-fala-ptbr-dataset.nearest_neighborsdata2plot_generated