datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
danbooru-tags-2024
danbooru-tags-2024
from datasets import load_dataset
ds = load_dataset(
"isek-ai/danbooru-tags-2024",
# revision="202412-at20250122", # optional
split="train",
)
Last updated: since 2005 to 2024/12/31, collected at 2025/01/22
openresearcher-sft-deep-research-cleaned
OpenResearcher SFT DeepResearch — Parquet Mirror
This is a re-hosted copy of the tool-reasoning SFT deep-research dataset by Aman Priyanshu, itself a cleaned/restructured version of the OpenResearcher Dataset from TIGER-AI-Lab.
Why this repo exists: the source wasn't laid out as ready-to-download Parquet files. This mirror simply stores the data as plain seed_*.parquet files so you can grab the whole dataset or a single segment easily. No changes were made to the content — all… See the full description on the dataset page: https://huggingface.co/datasets/DanielTobi0/openresearcher-sft-deep-research-cleaned.danbooru-wiki-2026
danbooru-wiki-2026-04-28
About
Wiki pages about the danbooru tags on danbooru.donmai.us. The wiki contains the description of each tag.
This dataset was collected from the wiki data using danbooru API and then merged with tags data (category_name & post_count). No entries were removed, no matter how many posts they have. The resulting set was manually filtered. Originally some of the lists, tag_group:, about:, help:, howto:, etc, pages were labled as "general"… See the full description on the dataset page: https://huggingface.co/datasets/kierarkia/danbooru-wiki-2026.danish-tool-dialogues-v9
danish-tool-dialogues-v1
Danish multi-turn tool-use conversations with reasoning, translated from the
Glaive subset of
Nanbeige/ToolMind
(Apache-2.0) by scripts/translate_toolmind_da.py.
Complements danish-tool-calls-v1, which is single-turn and synthetic. Here
the conversations run several turns, tool results are fed back, and the
assistant reasons before calling.
split
rows
train
34,168
eval_seen_tools
698
eval_unseen_tools
768
eval_seen_sym
752… See the full description on the dataset page: https://huggingface.co/datasets/jensjepsen/danish-tool-dialogues-v9.danbooru-tags-2023
danbooru-tags-2023
A dataset of danbooru tags.
Dataset information
Generated using danbooru and safebooru API.
The dataset was created with the following conditions:
Subset name
all
safe
API Endpoint
https://danbooru.donmai.us
https://safebooru.donmai.us
Date
2005-01-01..2023-12-31
2005-01-01..2023-12-31
Score
>0
>0
Rating
g,s,q,e
g
Filetype
png,jpg,webppng,jpg,webp
Size (number of rows)
6,574,149
1,387,371
Usage
pip install datasets… See the full description on the dataset page: https://huggingface.co/datasets/isek-ai/danbooru-tags-2023.danish-tool-dialogues-v6
danish-tool-dialogues-v1
Danish multi-turn tool-use conversations with reasoning, translated from the
Glaive subset of
Nanbeige/ToolMind
(Apache-2.0) by scripts/translate_toolmind_da.py.
Complements danish-tool-calls-v1, which is single-turn and synthetic. Here
the conversations run several turns, tool results are fed back, and the
assistant reasons before calling.
split
rows
train
17,667
eval_seen_tools
722
eval_unseen_tools
779
933 distinct tools; 59 names… See the full description on the dataset page: https://huggingface.co/datasets/jensjepsen/danish-tool-dialogues-v6.danbooru-tags-2024
danbooru-tags-2024
from datasets import load_dataset
ds = load_dataset(
"isek-ai/danbooru-tags-2024",
# revision="202412-at20250122", # optional
split="train",
)
Last updated: since 2005 to 2024/12/31, collected at 2025/01/22
danish-tool-dialogues-v7
danish-tool-dialogues-v1
Danish multi-turn tool-use conversations with reasoning, translated from the
Glaive subset of
Nanbeige/ToolMind
(Apache-2.0) by scripts/translate_toolmind_da.py.
Complements danish-tool-calls-v1, which is single-turn and synthetic. Here
the conversations run several turns, tool results are fed back, and the
assistant reasons before calling.
split
rows
train
17,160
eval_seen_tools
701
eval_unseen_tools
768
925 distinct tools; 59 names… See the full description on the dataset page: https://huggingface.co/datasets/jensjepsen/danish-tool-dialogues-v7.danish-tool-dialogues-v4
danish-tool-dialogues-v1
Danish multi-turn tool-use conversations with reasoning, translated from the
Glaive subset of
Nanbeige/ToolMind
(Apache-2.0) by scripts/translate_toolmind_da.py.
Complements danish-tool-calls-v1, which is single-turn and synthetic. Here
the conversations run several turns, tool results are fed back, and the
assistant reasons before calling.
split
rows
train
17,598
eval_seen_tools
762
eval_unseen_tools
772
932 distinct tools; 59 names… See the full description on the dataset page: https://huggingface.co/datasets/jensjepsen/danish-tool-dialogues-v4.danbooru-tags-2016-2023
danbooru-tags-2016-2023
A dataset of danbooru tags.
Dataset information
Generated using danbooru and safebooru API.
The dataset was created with the following conditions:
Subset name
all
safe
API Endpoint
https://danbooru.donmai.us
https://safebooru.donmai.us
Date
2016-01-01..2023-12-31
2016-01-01..2023-12-31
Score
>0
>0
Rating
g,s,q,e
g
Filetype
png,jpg,webppng,jpg,webp
Size (number of rows)
4,601,557
1,186,490
Usage
pip install… See the full description on the dataset page: https://huggingface.co/datasets/isek-ai/danbooru-tags-2016-2023.danish-tool-dialogues-v5
danish-tool-dialogues-v1
Danish multi-turn tool-use conversations with reasoning, translated from the
Glaive subset of
Nanbeige/ToolMind
(Apache-2.0) by scripts/translate_toolmind_da.py.
Complements danish-tool-calls-v1, which is single-turn and synthetic. Here
the conversations run several turns, tool results are fed back, and the
assistant reasons before calling.
split
rows
train
17,596
eval_seen_tools
762
eval_unseen_tools
772
932 distinct tools; 59 names… See the full description on the dataset page: https://huggingface.co/datasets/jensjepsen/danish-tool-dialogues-v5.danish-icl-schema-format-v3
danish-icl-json-v3
In-context-learning rows derived from jensjepsen/danish-json-grpo-v1. Each row
packs 1-5 worked examples into a single user turn, followed by a held-out
passage; the assistant turn is the answer for that passage. No instruction is
included, so both the schema and the output format have to be inferred from the
examples. Two axes vary per row and are held constant within a row: the schema
(134 field-sets) and the output format (10 renderers — JSON, key: value… See the full description on the dataset page: https://huggingface.co/datasets/jensjepsen/danish-icl-schema-format-v3.danish-ner-sft-v1
danish-ner-sft-v1
Danish named-entity SFT rows derived from chcaa/dansk-ner
(DANSK), which annotates Danish Gigaword text with the 18 OntoNotes entity
types using the same scheme in every split. Each row asks for a random subset
of entity types (1-6, including types absent from the passage so the empty
marker is exercised) and renders the answer in one of fourteen output formats.
Ten are key-value, listing the extracted entities: JSON, key: value,
key=value, [key] value, value… See the full description on the dataset page: https://huggingface.co/datasets/jensjepsen/danish-ner-sft-v1.danish-tool-dialogues-v8
danish-tool-dialogues-v1
Danish multi-turn tool-use conversations with reasoning, translated from the
Glaive subset of
Nanbeige/ToolMind
(Apache-2.0) by scripts/translate_toolmind_da.py.
Complements danish-tool-calls-v1, which is single-turn and synthetic. Here
the conversations run several turns, tool results are fed back, and the
assistant reasons before calling.
split
rows
train
34,168
eval_seen_tools
698
eval_unseen_tools
768
eval_seen_sym
752… See the full description on the dataset page: https://huggingface.co/datasets/jensjepsen/danish-tool-dialogues-v8.danish-tool-dialogues-v3
danish-tool-dialogues-v1
Danish multi-turn tool-use conversations with reasoning, translated from the
Glaive subset of
Nanbeige/ToolMind
(Apache-2.0) by scripts/translate_toolmind_da.py.
Complements danish-tool-calls-v1, which is single-turn and synthetic. Here
the conversations run several turns, tool results are fed back, and the
assistant reasons before calling.
split
rows
train
17,354
eval_seen_tools
756
eval_unseen_tools
768
908 distinct tools; 58 names… See the full description on the dataset page: https://huggingface.co/datasets/jensjepsen/danish-tool-dialogues-v3.danish-json-grpo-v1
danish-json-grpo-v1
10,015 Danish prompts for schema-directed JSON generation, built for GRPO training with a deterministic verifier (parse + key-set match + optional grounding penalty).
Task types
task_type
share
shape
extract
42%
Danish passage + schema → JSON grounded in passage
generate
26%
"Give me JSON for X with fields Y" (values open-ended)
rewrite
22%
Bullet list / semicolon-separated data → JSON with same info
fill_template
10%
JSON… See the full description on the dataset page: https://huggingface.co/datasets/jensjepsen/danish-json-grpo-v1.Barcenas-Personas-Mexico
Barcenas Personas México
Dataset sintético sociodemográfico de mayor granularidad geográfica publicado para México: 1,000,000 de personas, 2,478 municipios y 428,041 hogares.
Un millón de personas sintéticas, estadísticamente calibradas contra fuentes oficiales (INEGI, ENOE, CONEVAL) y distribuidas en los 2,478 municipios de las 32 entidades federativas de México. Cada persona cuenta con una ficha demográfica completa y 8 facetas narrativas en español (~1,100 palabras por… See the full description on the dataset page: https://huggingface.co/datasets/Danielbrdz/Barcenas-Personas-Mexico.danish-tool-dialogues-v1
danish-tool-dialogues-v1
Danish multi-turn tool-use conversations with reasoning, translated from the
Glaive subset of
Nanbeige/ToolMind
(Apache-2.0) by scripts/translate_toolmind_da.py.
Complements danish-tool-calls-v1, which is single-turn and synthetic. Here
the conversations run several turns, tool results are fed back, and the
assistant reasons before calling.
split
rows
train
17,138
eval_seen_tools
740
eval_unseen_tools
769
903 distinct tools; 57 names… See the full description on the dataset page: https://huggingface.co/datasets/jensjepsen/danish-tool-dialogues-v1.danish-tool-dialogues-v2
danish-tool-dialogues-v1
Danish multi-turn tool-use conversations with reasoning, translated from the
Glaive subset of
Nanbeige/ToolMind
(Apache-2.0) by scripts/translate_toolmind_da.py.
Complements danish-tool-calls-v1, which is single-turn and synthetic. Here
the conversations run several turns, tool results are fed back, and the
assistant reasons before calling.
split
rows
train
17,354
eval_seen_tools
756
eval_unseen_tools
768
908 distinct tools; 58 names… See the full description on the dataset page: https://huggingface.co/datasets/jensjepsen/danish-tool-dialogues-v2.danish-textman-v2
danish-textman-v2
Danish text-manipulation tasks over Wikipedia passages. Supersedes
jensjepsen/danish-textman-v1, which is identical except
that it also contained a textman_extraction subtype.
split
rows
train
99,610
validation
4,997
subtype
rows
textman_summary
20,937
textman_genre_transform
20,934
textman_style_transfer
20,923
textman_rewrite
20,912
textman_elaborate
20,901
What changed
textman_extraction is removed. It… See the full description on the dataset page: https://huggingface.co/datasets/jensjepsen/danish-textman-v2.danbooru2023-metadata-database
Metadata Database for Danbooru2023
Danbooru 2023 datasets: https://huggingface.co/datasets/nyanko7/danbooru2023
The latest entry of this database is id 7,866,491. Which is newer than nyanko7's dataset.
This dataset contains a sqlite db file which have all the tags and posts metadata in it.
The Peewee ORM config file is provided too, plz check it for more information. (Especially on how I link posts and tags together)
The original data is from the official dump of the posts info.… See the full description on the dataset page: https://huggingface.co/datasets/KBlueLeaf/danbooru2023-metadata-database.Prompt-Separation
Prompt-Separation
Voice-typed podcast prompt transcripts decomposed into structured fields:
discrete prompts (asks), a list of context chunks, and free-form
host_notes.
The dataset supports training a small model that, given a single voice-typed
message, recovers the structured fields an AI host would consume — separating
"what is the user actually asking?" from "what is the surrounding context?"
from "how should the response be shaped?".
Source
Prompts come from the… See the full description on the dataset page: https://huggingface.co/datasets/danielrosehill/Prompt-Separation.castillo
🏰 CASTILLO: Characterizing Response Length Distributions in Large Language Models
The CASTILLO dataset is designed to support research on the variability of response lengths in large language models (LLMs). It provides statistical summaries of output lengths across 13 open-source LLMs evaluated on 7 instruction-following datasets. For each unique ⟨prompt, model⟩ pair, 10 independent responses were generated using fixed decoding parameters, and key statistics were recorded—such as… See the full description on the dataset page: https://huggingface.co/datasets/danfperam/castillo.danish-tool-worlds-v4-declared
danish-tool-dialogues-v1
Danish multi-turn tool-use conversations with reasoning, translated from the
Glaive subset of
Nanbeige/ToolMind
(Apache-2.0) by scripts/translate_toolmind_da.py.
Complements danish-tool-calls-v1, which is single-turn and synthetic. Here
the conversations run several turns, tool results are fed back, and the
assistant reasons before calling.
split
rows
train
25,444
eval_seen_tools
550
eval_unseen_tools
1,012
eval_seen_sym
560… See the full description on the dataset page: https://huggingface.co/datasets/jensjepsen/danish-tool-worlds-v4-declared.danish-icl-schema-format-v1
danish-icl-json-v1
In-context-learning rows derived from jensjepsen/danish-json-grpo-v1. Each
row packs 1-5 worked examples sharing a JSON schema into a single user turn,
followed by a held-out passage; the assistant turn is the answer for that
passage. No instruction is included, so the schema and the output format have
to be inferred from the examples. In roughly half the rows the field names are
replaced by meaning-free symbols (alfa/beta/..., kat_a/..., f1/...,
foo/bar/...)… See the full description on the dataset page: https://huggingface.co/datasets/jensjepsen/danish-icl-schema-format-v1.danish-icl-schema-format-v2
danish-icl-json-v2
In-context-learning rows derived from jensjepsen/danish-json-grpo-v1. Each row
packs 1-5 worked examples into a single user turn, followed by a held-out
passage; the assistant turn is the answer for that passage. No instruction is
included, so both the schema and the output format have to be inferred from the
examples. Two axes vary per row and are held constant within a row: the schema
(134 field-sets) and the output format (8 renderers — JSON, key: value… See the full description on the dataset page: https://huggingface.co/datasets/jensjepsen/danish-icl-schema-format-v2.qwen25-math-sft-long-15k-v1
Qwen2.5 Math SFT Long 15K v1
A reproducible Long CoT SFT dataset for Qwen2.5-Math-1.5B with max context 3,840 tokens (below 4,096 limit).
Source: open-r1/OpenR1-Math-220k
Total samples: 15,000
Decontaminated against GSM8K, SVAMP, MATH-500, AIME 2026.
MHProbes
MHProbes Transition Pairs
This dataset contains synthetic transition-level contrastive pairs for harmful assistant behavior research.
Contents
synthetic_transition_pairs.csv: one row per user_message -> target_bot_code transition, with positive and negative bot replies side by side.
Row Structure
Each row includes:
transition_id
pair_id
user_taxonomy_id
user_taxonomy_name
user_message
target_bot_taxonomy_id
target_bot_name
positive_generation_status… See the full description on the dataset page: https://huggingface.co/datasets/danielfein/MHProbes.danish-if-grpo-combined-v3
danish-if-grpo-combined-v3
10,000 Danish instruction-following prompts for GRPO reward training,
built by rewriting danish-instruction-following-v4 prompts with a
sampled mix of our 46 Danish constraints + ~24 Google IFEval-schema
constraints.
What's new vs v2
Structural fixes (see scripts/constraint_compat.py on
github):
0% impossible constraint combos (v2 had 15.2%; e.g. constrained_response
paired with 20+ content-shape rules that can't co-satisfy). Enforced… See the full description on the dataset page: https://huggingface.co/datasets/jensjepsen/danish-if-grpo-combined-v3.WildDelusionVerified
WildDelusionVerified
WildDelusionVerified is a precision-oriented collection of 433 contextualized
target turns from 232 source conversations that potentially endorse highly
implausible beliefs. The public target turns were mined from WildChat-4.8M-Full
and ShareChat. The mining pipeline also found three verified LMSYS-Chat-1M
targets, but their conversation text is excluded from this release under that
dataset's terms. The intended use is evaluating how language models
respond… See the full description on the dataset page: https://huggingface.co/datasets/danielfein/WildDelusionVerified.