YC–001Agentic work
GDPval-AA v2
Professional work-product data built from source-grounded briefs that end in editable documents, spreadsheets, slides, models, and decisions.
Synthetic data infrastructure for frontier models
Frontier Data at Scale.
Benchmark-shaped, training-ready data for agents and reasoning models that need to act, recover, explain, and verify—not merely imitate.
RenewableNew task families rather than paraphrase loops.
ExecutableReal environments, tools, documents, and state.
VerifiableHidden tests, rubrics, provenance, and private holdouts.
10 capability-shaped families
Dataset systems
Ten frontier data systems. One measurable contract.
Each family converts an evaluation-shaped capability into renewable original tasks, observable trajectories, critiques, repairs, and a sealed assessment boundary.
10 systems shown
YC–002Tool use
𝜏³-Banking
Stateful banking and customer-service simulations with policy constraints, hidden account state, multi-turn interaction, and exact action checks.
YC–003Agentic code
Terminal-Bench v2.1
Long-horizon terminal tasks in containers with real filesystem and service state, hidden tests, recovery branches, and complete execution traces.
YC–004Scientific code
SciCode
Research-code tasks spanning implementation, debugging, numerical experiments, and scientific libraries, with tests and result-level checks.
YC–005Broad reasoning
Humanity's Last Exam
Broad expert-level reasoning and knowledge tasks with adversarial distractors, auditable solutions, and carefully sealed answer keys.
YC–006Scientific reasoning
GPQA Diamond
Graduate-level science question families with difficult distractors, explanation scoring, and expert-reviewed provenance.
YC–007Physics
CritPt
Physics problem-solving and critique data focused on derivations, assumptions, units, counterexamples, and repairable reasoning traces.
YC–008Knowledge reliability
AA-Omniscience
Knowledge-reliability data balancing correct answers, calibrated abstention, source checks, uncertainty labels, and hallucination penalties.
YC–009Long context
AA-LCR
Long-context reasoning corpora with distributed evidence, contradiction tracking, retrieval traps, structured citations, and answer provenance.
YC–010Ulam mathematics
Ulam Math Data
AIME-style exact-answer RLVR, graduate and research Harbor environments, and open research problems for multi-rollout post-training.
Evidence boundary. Each detailed page labels supplied blind-run evidence, linked public samples, or a delivery blueprint. Yotta Content tasks are original clean-room systems, not official benchmark releases, replicas, or affiliations.
Production loop
Data that behaves like an evaluation.
Every task starts with a capability contract and ends with evidence that can survive execution, review, replay, and adversarial checking.
01 Target
Define the capability
Start from a failure surface, not a prompt genre. Specify what success changes in the world.
02 Generate
Build task families
Create structural variants with new state, evidence, tools, constraints, and adversarial branches.
03 Execute
Check every claim
Run code, rebuild worlds, inspect artifacts, grade criteria, and retain the full trajectory.
04 Release
Protect the boundary
Separate training packs, repair pairs, pilot sets, and private holdouts with traceable provenance.
The output is a renewable data product: tasks, environments, source packs, event traces, labels, verifier bundles, and a changelog.
Inspect every environmentQuality principles
Scale the frontier, not the noise.
Volume matters only after novelty, solvability, verifier coverage, and contamination boundaries are measurable.
Structural novelty over paraphrase volume
Change mechanisms, state, evidence, tool topology, and failure opportunities—not only wording.
Verification before confidence
Prefer executable assertions, reproducible artifacts, and criterion-level review to opaque judge scores.
Failures are first-class data
Preserve diagnosis, partial progress, critique, repair, and the earliest consequential mistake.
Private holdouts remain private
Separate prompts, source metadata, graders, and exposure logs from the training path.
Artifacts, not just answers
Evaluate the document, codebase, environment state, proof object, decision, or tool action the model leaves behind.
Journal
Field notes from the data frontier.
A reusable editorial layer for dataset releases, methodology notes, verifier audits, and benchmark-inspired research.
Building a Terminal-Bench v2.1–shaped data system
How to turn terminal capability into renewable worlds, state contracts, hidden checks, failure labels, and repair trajectories.
Read field noteThree tracks for mathematical post-training
Exact-answer AIME-style RLVR, graduate and research Harbor environments, and open problems for multi-rollout training.
Read field noteEvidence-backed pages for every benchmark family
See which programs include supplied blind runs and which describe the complete delivery structure without a model score.
Browse all pagesStart with a failure surface
Bring us the capability your current dataset cannot fix.
We provide high-quality off-the-shelf datasets, RL tasks, and custom data.