AfterQuery closes $30M Series A at $300M valuationAfterQuery closes $30M Series A at $300M valuation and surpasses $100M revenue run rate
Read blogDataset catalog
Browse AfterQuery's commercial training-data programs, custom dataset capabilities, and public research datasets for LLMs and AI agents.
Commercial availability and terms are confirmed per engagement. Public repositories link directly to their official source so you can review the current files, metadata, and license status.
These programs are named in AfterQuery's published research. Request current availability, scope, delivery format, and terms for your training objective.
Commercial · Request access
Professional-work tasks with file-grounded inputs, multi-step analysis, judged outputs, rubrics, and deliverables such as spreadsheets, documents, and reports.
Best for
SFT warmup and reinforcement learning for office, research, and professional-work agents.
Published evidence
NVIDIA used AfterQuery tasks for SFT warmup and pivot RL. Its published GDPval ablation rose from 35.3 without warmup to 46.7 with warmup.
Commercial · Request access
Multi-turn customer-service trajectories built around policies, APIs, tools, user simulation, and passing task outcomes across public and AfterQuery-created domains.
Best for
SFT and agent training for policy-following, tool selection, argument accuracy, and user coordination.
Published evidence
An AfterQuery experiment used 1,057 passing rollouts covering 500 unique tasks across six domain variants and reported gains of up to 4.33× on unseen official τ² test tasks.
Commercial · Request access
Successful terminal-agent trajectories and verifier-backed RL tasks spanning exploration, planning, editing, testing, debugging, and task completion in isolated environments.
Best for
SFT and RLVR for terminal agents, SWE agents, tool use, debugging, and long-horizon coding workflows.
Published evidence
AfterQuery reported improving gpt-oss-20b from 3.1% to 17.0% on Terminal-Bench 2.0, with zero overlap between the training tasks and the official evaluation set.
When an existing dataset does not match the target distribution, AfterQuery can scope a pilot around a defined capability, workflow, domain, rubric, or model failure.
Supervised fine-tuning demonstrations and reasoning traces
Preference, ranking, critique, and RLHF data
Rubric- and verifier-based reinforcement-learning tasks
Tool-calling, browser, and computer-use trajectories
Code, debugging, deep-research, and multimodal data
Custom evaluations and held-out benchmarks
01
Share the model capability, workflow, domain, and evaluation target.
02
Review a relevant commercial dataset or scope a custom pilot.
03
Confirm delivery schema, training rights, security terms, quality criteria, timeline, and pricing for that engagement.
04
Evaluate the pilot against the target capability before scaling production.
These repositories are publicly accessible. Check the linked source before production use: where no license is published, do not assume commercial reuse rights.
Financial question-answering examples built from primary financial documents, with context, answers, reasoning, source links, and question types.
Full-stack application-building tasks with product prompts and functional rubrics for evaluating coding agents and AI app builders.
Client-style briefs for evaluating the visual design quality of AI website builders, with a main set spanning websites and web apps across five categories.
Finance and spreadsheet tasks with context files, golden outputs, domain labels, and the tools required to complete each workflow.
A human-evaluated benchmark for vulnerability assessment, detection, explanation, remediation, and test-plan generation in real software repositories.
A read-only SQLite bundle that supports deterministic, synonym-aware grading for computational-biology agent tasks.