−15pp
Lower model accuracy on our tasks than on similar data from other providers.
15 pp
Ours
Others
Task data for frontier models & agents
Data frontier models haven’t solved yet.
Original, expert-built tasks for training and evaluating LLMs and agents.
Trusted by
(01) Thesis
Benchmarks saturate. Models get good at what they have already seen. We write the tasks they haven’t — original, verified, and hard enough to move the frontier.
(02) Evidence
Code is our core.
On Deep SWE-style tasks, we benchmarked our data against comparable datasets from other leading data providers. Frontier models struggle more and reason longer on ours.
+20rounds
Longer agent trajectories on our tasks, on average, than on comparable data.
Others
Ours
Internal comparison on Deep SWE-style task data. Accuracy measured with GPT-5.6-sol; trajectory length measured in agent interaction rounds.
(03) Capabilities
Four domains.
One standard.
Every dataset is built from scratch for a target capability — never scraped, never recycled.
-
01
Coding
Repository-scale software engineering, algorithmic and optimization problems.
Core strength -
02
Complex reasoning
Expert-level problems across mathematics, science and beyond.
-
03
Multimodal
Vision-language reasoning over figures, diagrams, charts and screens.
-
04
Agents
Long-horizon, multi-tool tasks that demand planning over many steps.
(04) Benchmarks
Built for the hardest evals.
We design original datasets around the capability requirements of today’s most demanding benchmarks — aligned in difficulty, never copied from them.
- 01ALE-BenchAlgorithmic engineering
- 02Frontier-EngEngineering
- 03HLEExpert reasoning
- 04HLE-VLMultimodal reasoning
- 05Frontier SWESoftware engineering
- 06Deep SWEAgentic coding
- 07Program BenchProgramming
- 08APEX AgentsLong-horizon agents
(05) About
Born in a research lab at Peking University.
Detamaning was founded by PhD researchers from the Institute of Software, School of Computer Science, Peking University. We bring research rigor to a single question: what data makes models genuinely more capable?
- Origin
- Peking University · School of CS
- Focus
- Task data for LLMs & agents
- Clients
- ByteDance · Alibaba · Tencent
- Based in
- Beijing, China
