datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
typed-decisions
Typed Decisions
A benchmark for typed probabilistic decisions over shared state. You give a
model one piece of unstructured state. It answers several typed questions about
that state at once, and every answer is a probability distribution rather than a
single label.
The schema follows the System One primitives used by
TypeSafe AI: noul, choice and score. A row replays against any API that implements that shape. This benchmark is
independent. It is not affiliated with TypeSafe… See the full description on the dataset page: https://huggingface.co/datasets/LocalLLaMA/typed-decisions.typed-decisions-causal-experimenttyped-decisions-v2
typed-decisions-v2
Corrected companion corpus to pngwn/typed-decisions
(the "v1" corpus) for the typed-decision baselines. v2 repairs the
synthetic-domain label/oracle inversion that was disclosed but not fixed in v1
(nanodiff REPORT.md, finding 6) and adds a raw-text dump so that any tokenizer
(GPT-2 and Qwen) can consume byte-identical examples.
The fix
Both defects live in the synthetic ticket-triage generator (code/build_dataset_v2.py,
applied to the v1… See the full description on the dataset page: https://huggingface.co/datasets/pngwn/typed-decisions-v2.tasksource-jev-typed-decisions
tasksource-jev-typed-decisions
One million human-labeled decisions from 500+ tasks, in one format for models
that read their answer criteria at runtime.
Most instruction data teaches a model to write. This dataset teaches it to
decide: given a state and a question, pick among the options it is handed,
rate on a scale it is handed, or give a calibrated probability. The options
change from row to row, so a model has to read them rather than memorize a label set.
Why… See the full description on the dataset page: https://huggingface.co/datasets/tasksource/tasksource-jev-typed-decisions.typed-decisions-synth
Typed Decisions Synth
This is the synthetic dataset I made for Hmm, a small open model that answers questions about your data with probabilities instead of text.
It has 7,414 cases with 25,859 questions across 149 domains and workflows. Every question has an answer and a soft label (a probability for every option), so you can train a model to be unsure when it should be.
Code and the model: github.com/n4ze3m/hmm
Note: Everything here is written and labelled by an LLM. Nobody… See the full description on the dataset page: https://huggingface.co/datasets/n4ze3m/typed-decisions-synth.typed-decisions
typed-decisions
A typed-decision corpus for training a masked-diffusion LM to emit calibrated
discrete decisions instead of text. Built for fine-tuning
Sebasdi/nanodiff-350m-base
(the LLaDA recipe).
The interface
Every example is a prompt plus a response, and every decision is a single
masked token. The answer is always one option letter A-J:
### State:
<unstructured state text>
### Question:
<the decision to make>
### Options:
A) yes
B) no
### Answer:
A
The… See the full description on the dataset page: https://huggingface.co/datasets/pngwn/typed-decisions.typed-decisions-v2-system-onetyped-decisions-code-holes
typed-decisions: code-holes
Single-token substitutions mined from the git history of 64 public repositories in the
kotoba-lang organisation, each turned into one choice question
with a gold answer: the token the commit actually put there. Part of
kotoba-lang/typed-decisions (README 第8反復),
whose trained model is com-kotobalabs/open-jev-deberta-v3-large.
Sibling dataset (ungoaled): com-kotobalabs/typed-decisions-repo-governance.
What a record is
state = the commit… See the full description on the dataset page: https://huggingface.co/datasets/com-kotobalabs/typed-decisions-code-holes.typed-decision-bench
Typed Decision Bench v0.3
Built by Blobfish AI. A benchmark for one-pass decision models: 5,387 items, 25 tasks, 5 use-case
suites. Blobfish designed the tasks, wrote the typed questions, framed each one as a decision a business actually
delegates (use case, vertical), drew stratified seeded panels, froze them, and built the scoring, the contamination
tiers and the quality scorecard. The underlying records are drawn from 21 openly licensed public datasets plus one
generator of… See the full description on the dataset page: https://huggingface.co/datasets/SamuelChien821/typed-decision-bench.typed-decisions-repo-governance
typed-decisions: repo-governance family
A seed, not a corpus (n=1 as of first publish, 2026-09-19): real governance findings from a
private software monorepo, each run through TypeSafe's Jev (a "System One"
model — it returns typed choice / score / noul decisions, never free text or code) as part of
kotoba-lang/typed-decisions, a reproduction and
measurement of the Jev shape whose trained model is published at
com-kotobalabs/open-jev-deberta-v3-large.
What this is… See the full description on the dataset page: https://huggingface.co/datasets/com-kotobalabs/typed-decisions-repo-governance.typed-decisions-pt-es
Typed Decisions PT/ES
Unofficial Brazilian Portuguese and Spanish translations of
LocalLLaMA/typed-decisions.
The dataset covers the choice, score, and noul typed-decision tasks. It preserves the
original case IDs, workflows, question IDs, task types, option keys, splits, gold labels, and
soft target distributions.
Contents
Config
Train
Test
pt
1,200
400
es
1,200
400
Each row contains:
id and workflow from the source case;
language, either pt or… See the full description on the dataset page: https://huggingface.co/datasets/telepatia-ai/typed-decisions-pt-es.notam-typed-decisions
NOTAM typed decisions
Frozen, hashed evaluation suites for typed decisions about NOTAMs — a choice, a
yes/no, or a score, each with a confidence — plus a catalogue of how NOTAMs describe
areas in free text. Built by Airside Labs so that any model,
served any way, can be scored on the same rows and read with the same per-class table
and calibration curve.
Not for operational use. These suites and the numbers quoted here are for research
and for triage tooling. NOTAMs are… See the full description on the dataset page: https://huggingface.co/datasets/AirsideLabs/notam-typed-decisions.mmlu_typed_decisionprocedural-typed-decisions
procedural-typed-decisions
Procedurally generated decision problems. Each row is one structured state
(JSON, or a table, CSV, key=value lines, or prose for the arithmetic,
retrieval, and aggregation configs) with several typed questions over that same state, following the
Jev / System One request shape: choice (pick one criterion), noul (a
number in [0, 1]; a probability or a yes/no), and score (an ordered rubric).
Every answer is computed exactly from the state by rules that… See the full description on the dataset page: https://huggingface.co/datasets/tasksource/procedural-typed-decisions.