datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
jev-bench
jev-bench
Real human-labeled data, reformatted into System One questions — with human label distributions wherever they exist.
22 configs · 166,054 rows · 22,773 test records · 4 calibration-gold configs · v0.1.1
Repo & engine · Source rationale · What we verified about Jev's API · Other independent Jev evaluations
jev-1.13.0 on every test record: crisp, grounded decisions land in the accurate-and-calibrated corner; ordinal ratings and anything humans disagree about do not.… See the full description on the dataset page: https://huggingface.co/datasets/Praveenrajus/jev-bench.DGUI_HYPERMEM-JEV
DGUI_HYPERMEM-JEV
The training "brain" for DGUI-HyperMem (DeckerGUI HyperMemory) — the self-hosted
memory MCP server. Every JEV reasoning decision the service makes is appended here as a
typed instruction row, so the corpus grows with real usage and can be used to fine-tune or
few-shot the JEV layer later.
Usage
from datasets import load_dataset
ds = load_dataset("ctaxnagomi/DGUI_HYPERMEM-JEV", split="train")
for row in ds.stream():
print(row["use_case"]… See the full description on the dataset page: https://huggingface.co/datasets/ctaxnagomi/DGUI_HYPERMEM-JEV.jev-distill-corpus
Jev-Gate Student B — LoRA Memory-Relevance Judge
A LoRA adapter (r=16, α=32, on q_proj/v_proj) on Qwen/Qwen2.5-0.5B-Instruct, distilled from the Jev typed-judgment API into a compact local judge for agent-memory gating.
What it does: given a query and a candidate memory passage, outputs P(relevant) as the calibrated yes probability read from the final-token logits of yes vs no. Used to filter which vector-recalled memories get injected into agent context (vector recall →… See the full description on the dataset page: https://huggingface.co/datasets/SargeDev/jev-distill-corpus.INSTRUCT_JEV
INSTRUCT_JEV
INSTRUCT_JEV is an instruction corpus built from the TypeSafe AI documentation
for Jev, the first System One model. It is structured around the three TypeSafe
question primitives - Choice, Noul and Score - and mirrors the raw corpus
captured in deckerGUI-jev_corpus_RAW.
Credits
INSTRUCT_JEV is a DeckerGUI project and exists because of the work below.
Who
Contribution
Link
TypeSafe AI
Jev - the first System One model - and the Choice / Noul… See the full description on the dataset page: https://huggingface.co/datasets/ctaxnagomi/INSTRUCT_JEV.mini-jev-runs
mini-Jev run records: 27 900 schema-driven decisions with full candidate logits
Every record is one decision a frozen Qwen/Qwen3-4B-Instruct-2507 made about one field of a JSON
schema that arrived with the request. The field was turned into a lettered multiple-choice question
(A = pay_bill, B = bill_balance, …), the model ran one forward pass, and the answer was
read from its next-token logits over the option letters. No token was generated.
The records keep what such a run… See the full description on the dataset page: https://huggingface.co/datasets/Mikhail/mini-jev-runs.JevForge-Mind2Web
JevForge Mind2Web Gold Decisions
Private research snapshot of gold candidate-decision records used by
JevForge.
What is included
Each JSONL record contains a page state, a choice question over candidate
elements, a noul question about one candidate, complete gold target
distributions, a website group, and source annotation metadata.
Split
Records
Websites
train
4,642
49
dev
786
6
calibration
400
6
test
800
8
ood
386
4
The 7,014 record IDs… See the full description on the dataset page: https://huggingface.co/datasets/AndeyTait/JevForge-Mind2Web.jev-distill-corpus-v3
Jev Distill Corpus v3
A 740,957-row calibrated typed-decision corpus in the TypeSafe System One schema
(noul / choice / score). Built for training small local "System One" judges
(fine-tuned BERT-class encoders or small Qwens) that read a state and a typed
question and return a calibrated probability distribution in one forward pass.
Streams
stream
rows
origin
yuri_v3
498,010
Synthetic operational scenarios across 53 domains (business, technical, agent… See the full description on the dataset page: https://huggingface.co/datasets/SargeDev/jev-distill-corpus-v3.jevlite_dataset
jevlite — synthetic decision questions with soft labels
5,866 typed questions about 978 synthetic program states — support tickets, SIEM alerts,
vendor invoices, AI agent transcripts, code reviews, incident logs, chat threads and job
applications — each answered by a teacher model as a full probability distribution
rather than a single label.
This is the synthetic portion of the training data for
vagmi/jev-lite, a decision model that reads a state
and a typed question and… See the full description on the dataset page: https://huggingface.co/datasets/vagmi/jevlite_dataset.jev-my-bro-dataset
jev-my-bro Governance Dataset
Provenance-aware English/Thai dataset for training and evaluating the typed
jev-my-bro decision model. Each case asks four structured governance
questions about an operation:
action: execute, ask_user, or reject
needs_review: whether explicit human review/approval is required
prohibited: whether the operation should be prohibited
risk: five-level operational risk
This snapshot contains 8,508 cases / 34,032 typed decisions.
This is… See the full description on the dataset page: https://huggingface.co/datasets/JonusNattapong/jev-my-bro-dataset.postflop-solver-reasoning-traces-1m
Postflop-Solver Reasoning Traces (1M, v2)
Teacher-forced chain-of-thought reasoning traces for Heads-Up No-Limit Texas
Hold'em postflop decisions, distilled from a GTO solver (postflop-solver)
plus a strong LLM teacher.
Each example pairs a game scenario with the known-optimal solver action and
a step-by-step natural-language justification of why that action is correct.
The teacher is conditioned on the gold action (teacher forcing), so every trace
supports the correct move —… See the full description on the dataset page: https://huggingface.co/datasets/jevonmao/postflop-solver-reasoning-traces-1m.jev-legal-judgment-tests
Jev-like legal judgment test sets
Two case-group-held-out evaluation splits: original_test (173 rows) and fresh_test (361 rows). They do not overlap the 1269-row training or 158-row validation data. Original test has 149 accepted rows; fresh test has 322 accepted rows. Filter audit_status == 'accepted' for the primary audited subset. Other rows retain review/conflict/instability flags.
Each row contains facts, a legal proposition, finite candidate labels, optional self-contained… See the full description on the dataset page: https://huggingface.co/datasets/Nebulaw1/jev-legal-judgment-tests.NoLimitHUPokernext-jev-laya-test
next-jev Laya decision test set
24182 questions. This is the project decision test set.
It is a frozen slice of public datasets, built with the same seeds and cuts as the
Laya benchmark runs. It is not a three-way NLI training split.
The bundle is CC BY-NC 4.0 because it contains CC BY-NC sources
(lmsys/toxic-chat, Tobi-Bueck/customer-support-tickets).
MS MARCO passages stay under Microsoft's non-commercial research terms.
BoolQ is CC BY-SA 3.0. Other rows keep the source… See the full description on the dataset page: https://huggingface.co/datasets/JonesLin/next-jev-laya-test.