datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
latentsKnowledge_distilled_dataset_by_NAGISA_V3
NAGISA_V3 teacher shards
Training data distilled from self-play of the search engine attic reading
the NNUE weights NAGISA_V3, in the shape the trainers read directly.
Starting from a balanced-opening book, the engine played itself at MultiPV=5,
recording the root score and the candidate moves at every ply;
manaka-teacher turned that corpus into raw MPK1 streams, and manaka-pack
folded identical positions into one row each and wrote these parquet
shards.
Rows: 280,621,202… See the full description on the dataset page: https://huggingface.co/datasets/qleap/Knowledge_distilled_dataset_by_NAGISA_V3.so-101_dataset02_20260827_130357This dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v3.0",
"fps": 30,
"features": {
"action": {
"dtype": "float32",
"shape": [
6
],
"names": [
"shoulder_pan.pos",
"shoulder_lift.pos",
"elbow_flex.pos",
"wrist_flex.pos",
"wrist_roll.pos",
"gripper.pos"… See the full description on the dataset page: https://huggingface.co/datasets/nagaenaga/so-101_dataset02_20260827_130357.SO-101_dataset03_20260827_145444This dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v3.0",
"fps": 30,
"features": {
"action": {
"dtype": "float32",
"shape": [
6
],
"names": [
"shoulder_pan.pos",
"shoulder_lift.pos",
"elbow_flex.pos",
"wrist_flex.pos",
"wrist_roll.pos",
"gripper.pos"… See the full description on the dataset page: https://huggingface.co/datasets/nagaenaga/SO-101_dataset03_20260827_145444.saccade-egomotion-bench
Saccade ego-motion benchmark
The stream, the raw decision signals, and the per-frame measurements behind
Saccade — an always-on edge VLM that re-encodes only
the image patches whose change ego-motion cannot explain.
This dataset exists so the central claim can be checked without running our code.
💻 Code: https://github.com/NagaYu/saccade
🤖 Model: https://huggingface.co/NagaYu/saccade-predictor
🚀 Demo: https://huggingface.co/spaces/NagaYu/saccade
The claim, in… See the full description on the dataset page: https://huggingface.co/datasets/NagaYu/saccade-egomotion-bench.Training_dataset_by_NAGISA_V4
NAGISA_V4 ply-37 teacher shards, games played to the end
Self-play of attic-gensfen reading the NNUE weights NAGISA_V4
(HalfKA-2304), in the shape the trainers read directly. Every game starts from a
balanced ply-37 position, makes no random moves, and runs until it actually
ends. Identical positions are folded into one row each.
67,108,864 rows — exactly 2^26
16 shards of 4,194,304 rows, 512 row groups each, zstd, 5,208,620,326 B total
Against
Opening_dataset_by_NAGISA_V4… See the full description on the dataset page: https://huggingface.co/datasets/qleap/Training_dataset_by_NAGISA_V4.Training_dataset_qsearched_by_NAGISA_V4
NAGISA_V4 teacher shards, moved to their quiescence leaves
Every record of
Training_dataset_by_NAGISA_V4
walked to the end of its quiescence variation, the deep search's value kept
there, and a policy fitted at the leaf itself.
The parent's positions are as its games reached them, with no quiescence
search — a row can sit in the middle of an exchange, where the evaluation
swings by a piece depending on whose turn it is to recapture. A value fitted on
those learns the swing. This… See the full description on the dataset page: https://huggingface.co/datasets/qleap/Training_dataset_qsearched_by_NAGISA_V4.Opening_dataset_by_NAGISA_V4
Opening dataset by NAGISA_V4
平手から 15手目までを方策ネット b15c256 で広げ、NNUE エンジン attic-gensfen
が NAGISA_V4 (HalfKA-2304) を読んで depth 9・MultiPV 5 で採点した序盤局面集。
manaka-teacher が MPK1 ストリームに落とし、manaka-pack が同一局面を 1 行に
畳み込んで、この parquet を書いた。学習側がそのまま読む形である。
自己対局のコーパスではない。 対局は 1 局も指していない。木を広げて採点しただけで、
局面同士に前後関係は無い。勝敗を持つデータが要るなら
qleap/Knowledge_distilled_dataset_by_NAGISA_V4
のほう。あちらと同じスキーマ・同じ変換規約なので、混ぜて読める。
行数: 19,896,088 (畳み込み前のレコード数: 19,898,611)
シャード数: 4 (data/teacher-00000.parquet …)
合計: 1,137… See the full description on the dataset page: https://huggingface.co/datasets/qleap/Opening_dataset_by_NAGISA_V4.Knowledge_distilled_dataset_by_NAGISA_V4
NAGISA_V4 teacher shards
Training data distilled from self-play of the generator attic-gensfen
reading the NNUE weights NAGISA_V4 (HalfKA-2304), in the shape the trainers
read directly. The engine played itself at depth 9 and MultiPV 5, recording the
root score and the candidate moves at every ply; manaka-teacher turned that
corpus into raw MPK1 streams, and manaka-pack folded identical positions
into one row each and wrote these parquet shards.
Rows: 1,221,920,925 (records… See the full description on the dataset page: https://huggingface.co/datasets/qleap/Knowledge_distilled_dataset_by_NAGISA_V4.isotope-bench
Isotope Bench
An indirect-prompt-injection benchmark for tool-calling agents, plus the
complete audit trail of one recorded run: 438 influence certificates, one for
every action an agent attempted across five defence conditions.
Built for Isotope, which tracks
untrusted influence inside the forward pass. The corpus is independent of that
method and usable with any defence.
💻 Code: https://github.com/NagaYu/isotope
🤗 Demo: https://huggingface.co/spaces/NagaYu/isotope
🤗… See the full description on the dataset page: https://huggingface.co/datasets/NagaYu/isotope-bench.wikisplit-pp
WikiSplit++
This dataset is the HuggingFace version of WikiSplit++.WikiSplit++ enhances the original WikiSplit by applying two techniques: filtering through NLI classification and sentence-order reversing, which help to remove noise and reduce hallucinations compared to the original WikiSplit.The preprocessed WikiSplit dataset that formed the basis for this can be found here.
Usage
import datasets as ds
dataset: ds.DatasetDict =… See the full description on the dataset page: https://huggingface.co/datasets/cl-nagoya/wikisplit-pp.deference-keigo-corpus
Deference — Japanese honorific (keigo) error corpus
A corpus for detecting and correcting errors in Japanese honorifics, constructed
mechanically from the norm rather than collected or generated by a model.
The classes, forms and conditions set out in the Council for Cultural Affairs'
report Keigo no Shishin (敬語の指針, 2007) are implemented as rules; correct
sentences are generated from those rules, and documented error types are then
injected — also by rule.
No LLM was… See the full description on the dataset page: https://huggingface.co/datasets/NagaYu/deference-keigo-corpus.downstep-bench
Downstep compound pitch-accent benchmark
Japanese noun compounds, their mora segmentation, and every accent their source
dictionaries attest. Built so that "the model has never seen this compound" is a
condition you can actually turn on, rather than a claim you have to trust.
No human annotation is present in this release. Every accent label here is
dictionary-derived. docs/ANNOTATION_GUIDELINES.md ships the protocol, the CSV
format and the agreement statistics for collecting… See the full description on the dataset page: https://huggingface.co/datasets/NagaYu/downstep-bench.rebar-structure
🧱 Rebar Structure
A corpus for restoring the heading hierarchy of Japanese documents from flat text and
for evaluating structure-aware chunking. Each record is a flattened document, its
gold heading tree (positions + depths), and a damaged variant simulating PDF/text
extraction.
Code: https://github.com/NagaYu/rebar
Model: https://huggingface.co/NagaYu/rebar-heading-classifier
Demo (Space): https://huggingface.co/spaces/NagaYu/rebar
Why it exists
The same… See the full description on the dataset page: https://huggingface.co/datasets/NagaYu/rebar-structure.kvine-agent-trajectory-bench
KVine agent-trajectory benchmark & resident-set policy training data 🌿
Two things from the KVine research prototype
(branch-aware KV cache for agent trajectories):
policy_train — self-supervised training data for the resident-set policy:
structural features of branches in KVine's trajectory tree, labelled with
whether the branch was revisited within the next 5 steps.
benchmark_results.json — the full A / B / C benchmark output (cumulative
prefill FLOPs, TTFT, recomputed tokens… See the full description on the dataset page: https://huggingface.co/datasets/NagaYu/kvine-agent-trajectory-bench.auto-wiki-nli-reward
AutoWikiNLI reward
A dataset constructed by generating hypothesis sentences corresponding to entailment and contradiction from Wikipedia text using Nemotron-4 340B.
Helpfulness and other scores are assigned using the Nemotron-4 340B reward model.
crowd-anonymity-sets
Crowd — anonymity-set sizes for attribute combinations
Read this first
This dataset describes nobody. Every row is generated from a probability
model over attribute values built from published aggregate statistics. There
is no person in it, no record to link, and no index that could be searched for
an individual. The prose is synthetic.
The label is a head-count, not an identity. Each row's target is
log10(number of people in the reference population matching… See the full description on the dataset page: https://huggingface.co/datasets/NagaYu/crowd-anonymity-sets.assay-receipts
Assay Receipt Corpus
Signed internal receipts from an inference provider that is sometimes cheating, together
with the verdict an auditor reached on each one and the ground truth of which model actually
served the request.
Each row is a real receipt, not a summary statistic: it carries the prompt and output token
ids, the JL-projected sketch of the provider's hidden_states, the sign/rank invariants, and
an HMAC signature. With the gpt2 weights you can recompute the sketch… See the full description on the dataset page: https://huggingface.co/datasets/NagaYu/assay-receipts.bleep-spans
Bleep spans — synthetic sensitive-speech regions with frame-accurate labels
Where sensitive information is spoken, and what kind it is — never what was
said.
Every recording is synthetic. No real telephone call, clinical recording, or any
other real speech was used, recorded, or derived from at any stage.
🤗 Model: NagaYu/bleep-0.09b
🎛️ Demo: NagaYu/bleep
What a row contains
utt_id, voice_key, condition, duration, subsets, and three parallel
arrays —… See the full description on the dataset page: https://huggingface.co/datasets/NagaYu/bleep-spans.Balanced_extended_dataset_by_NAGISA_V4
NAGISA_V4 balanced-extended teacher data
互角局面集を 37手目まで広げた将棋の局面に、NNUE エンジン (NAGISA_V4,
HalfKA-2304) で depth 9・MultiPV 5 の評価値と候補手を付けた教師データ。
自己対局のコーパスではない。 対局は 1 局も指していない。1 行が 1 局面で、
局面同士に前後関係は無い。勝敗を持つデータが要るなら
qleap/Knowledge_distilled_dataset_by_NAGISA_V4
のほう。
局面数: 25,075,766
シャード数: 26 (data/shard-00000.parquet … data/shard-00025.parquet)
合計: 1,606,525,370 B
from datasets import load_dataset
ds = load_dataset("qleap/Balanced_extended_dataset_by_NAGISA_V4", split="train")… See the full description on the dataset page: https://huggingface.co/datasets/qleap/Balanced_extended_dataset_by_NAGISA_V4.molt-benchmark-results
Molt · elastic on-device inference measurements
Everything measured while building Molt, a
runtime that moves a running generation onto a smaller model between two
tokens, carrying the KV cache across, so an on-device LLM under memory pressure
is neither reclaimed by the OS nor restarted from the prompt.
Published so the claims can be checked rather than taken on trust. The figures in
the repo README and the results page are generated from these files; nothing is
transcribed by… See the full description on the dataset page: https://huggingface.co/datasets/NagaYu/molt-benchmark-results.sludge-ui-counterfactuals
Sludge counterfactual UI corpus
This model does not determine legality.
It reports provisions that may be implicated and the screen elements that are the
factual basis for looking at them. Whether a provision is actually engaged depends on facts
no UI tree contains — the purposes of processing, the legal basis relied on, the audience, the
rest of the journey, prior consent, sector rules — and is an assessment for a qualified human.
It has no feature that labels a… See the full description on the dataset page: https://huggingface.co/datasets/NagaYu/sludge-ui-counterfactuals.clearance-bench
Clearance Benchmark
A fully synthetic enterprise corpus for measuring how well a retrieval system
contains information flow: heavily overlapping ACLs, sensitivity levels,
time-limited grants, a grant/revoke timeline, and evaluation queries.
Generated by clearance.synth with seed 7. No real
documents, no real access-control lists, and no real identities. Regenerating
with the same seed reproduces this dataset byte for byte.
Why this exists
Query-time ACL filtering… See the full description on the dataset page: https://huggingface.co/datasets/NagaYu/clearance-bench.halfword-bench
Halfword benchmark: conversational text under access-method cost models
This dataset pairs public conversational sentences with timing cost models for AAC
access methods, so that a prediction system can be scored in seconds to utterance
rather than in keystrokes saved.
It contains no data from AAC users. It is public conversational text plus simulation.
Configurations
utterances (12565 rows) -- normalised sentences with history,
pseudo-speaker, source and that… See the full description on the dataset page: https://huggingface.co/datasets/NagaYu/halfword-bench.claimcheck-eval
ClaimCheck Evaluation Set
158 hand-checked cases for claim-level groundedness verification: given an
answer and the context that was supplied to the model, is each specific claim in
that answer supported, derived, approximate, unsupported or contradicted?
Bilingual (87 English / 71 Japanese). Every case was run through
ClaimCheck v1.1.0, and the
dataset records what the tool actually returned, not just what it should
have returned.
Cases
158
Languages
English 87 ·… See the full description on the dataset page: https://huggingface.co/datasets/NagaYu/claimcheck-eval.Z3-Verified-Reasoning-Graphs
Z3-Verified Constraint Reasoning Dataset
5k Baseline · Production-Ready · Zero Label Noise
The Problem This Solves
Most synthetic reasoning datasets only show the "happy path". Real reasoning requires knowing when to backtrack.
Open-source LLMs hallucinate on constraint satisfaction problems because they are trained on fluent-sounding but logically inconsistent traces. This dataset is different:
❌ No LLM-generated reasoning — zero hallucinations, zero label noise
✅… See the full description on the dataset page: https://huggingface.co/datasets/nagygabor/Z3-Verified-Reasoning-Graphs.scribe-koyobun-usage
Scribe usage-judgment dataset
Span-level data for judging context-dependent kanji/kana usage in Japanese official writing.
Generated by scripts/build_dataset.py in the GitHub repo.
What the claim rests on
The center of this dataset is the hard split: occurrences of words that appear in both usages
(kana and kanji) across the corpus. Because uniform dictionary replacement collapses a word to a single
spelling, it is structurally forced to mislabel one side of this… See the full description on the dataset page: https://huggingface.co/datasets/NagaYu/scribe-koyobun-usage.litmus-kernels
Litmus Kernel Verification Corpus
Correct and deliberately-broken Triton kernels, each broken one shipped with
the input that exposes it.
The corpus exists to measure one thing: how much of what a fixed-shape
torch.rand() allclose test calls "correct" actually is. On this corpus the
answer is that 88% of the planted bugs pass that test.
Columns
column
meaning
name
kernel identifier
family
elementwise / reduction / softmax / layernorm / matmul /… See the full description on the dataset page: https://huggingface.co/datasets/NagaYu/litmus-kernels.amazon-laptop-product-catalogdendro-lowbackground
Dendro low-background corpus
5034 arXiv records annotated with archival evidence of when they existed, produced by
Dendro v0.1.0.
2500 of them (49.7%) are low-background: an independent registration record
places them before 2021-01-01, i.e. before large-scale text generation. The name is from
metallurgy — low-background steel is steel smelted before the 1945 atmospheric tests: not
special steel, just ordinary steel that happens to predate the contamination, and valuable
because… See the full description on the dataset page: https://huggingface.co/datasets/NagaYu/dendro-lowbackground.