Large-scale corpora
6B+text records
- Private code repos
- 15K
- Speech, 26 languages
- 890K hrs
- Aligned A/V & subtitles
- 105K hrs
- Clinical imaging studies
- 4.35M
Frontier data lab · Silicon Valley
A frontier data lab
in Silicon Valley
IcarusTech runs on an expert network. We supply training data and RL environments across pretraining, midtraining, post-training and world models. Tasks are written by working experts, difficulty is measured on current frontier models, and grading is executable and reproducible.
30+
Industry partners
Real business settings, industry data and domain experts
200+
Academic PhDs
Research-grade problems, rubrics and independent review
50K
Annotation workforce
Collection, annotation and QA at scale
Data products Four stages · one supplier
Data for every training stage
Organised by training stage: large-scale corpora for pretraining, agent trajectories for midtraining, RL environments and expert data for post-training, and embodied and simulation data for world models.
Knowledge and language1PretrainingStage 01
How agents behave2MidtrainingStage 02
Verifiable capability gains3Post-trainingStage 03
Understanding the physical world4World modelsStage 04
Real-user agent trajectories
7.85Msessions
- Avg. messages per session
- 119.7
- Long-horizon, 100+ msgs
- 2.72M
- Claude family · GPT family
- 49% · 51%
- Coding-agent runs, pass & fail
- 505K
RL environments & expert data
200K+RL tasks
- Agents' Last Exam · 10 lines
- 970
- Terminal-Bench 3 & 4
- 553
- Terminal-Bench Science
- 1,218 + 136
- Game, mobile & robotics envs
- 5,000+
Embodied & simulation
355Kepisodes
- World-model sim, multi-view
- 10K hrs
- Robot manipulation RL envs
- 5,000+
- Original filmed footage
- 1,500 hrs
- Game source & assets
- 65 titles
Figures are approximate and reflect stock or standing capacity. Post-training centres on Agents' Last Exam, Terminal-Bench 3 & 4 and Terminal-Bench Science. Off the shelf · Built to order
Pretraining Stage 01
Large-scale pretraining corpora
Every dataset carries its source, licence and screening record, and is kept separate by language, source and platform rather than pooled.
Text · news, forums & public opinion
6B+records
Platform and print archives with outlet, date and section kept; each language stored separately.
- Chinese platforms (3 archives)
- 4.7B
- Overseas forums (2 archives)
- 1.5B
- Overseas news · EN / VI / TH
- 234M
Academic, question banks & professional
300M+records
Figures and references intact; question banks kept in source form, by discipline.
- Papers and theses
- 150M
- Question banks, K-12 to professional
- 150M
- Filings and research reports
- 22M
Code
15Kprivate repos
Never published, with full git history; screened for credentials, personal data and malware, with overlap against public code measured.
- Lines of code
- 1.6B
- Tokens
- 10B
- Quant strategy source
- 200K
Speech & audio
890Khours
Each language and each dialect is its own corpus, never pooled into one multilingual set.
- Languages and varieties
- 26
- Dialects & low-resource
- 13K hrs
- Music with lyric files
- 3.7M tracks
Video
105Khours
Audio, video and subtitle tracks aligned across 9 languages, with translation tracks where they exist.
- Largest language
- 50K hrs
- Original filmed footage
- 1,500 hrs
- Scripted action capture
- 150 hrs / mo
Image
4.35Mclinical studies
Real hospital imaging examinations, with structured findings.
- Planning & design material
- 130 TB
- Mobile UI screenshots
- 200K
- Artworks with artist notes
- 200K
Midtraining · real-user trajectories Stage 02
Real-user agent trajectories
Complete sessions of real users getting work done with frontier models, successes and failures alike. Normalised into one format across models, ready to use as a single midtraining corpus.
Up to 100 messagesOver 100: long-horizon
<0.1%2–4
7.4%5–10
16.8%11–20
23.9%21–50
17.2%51–100
16.8%101–200
14.1%201–500
3.7%501+
Claude family 48.5% · 3.81MGPT family 51.5% · 4.04M
AlsoCoding-agent trajectories 505K, pass and fail kept, replayableComputer-use trajectories 21K, step by stepBimanual teleop demos to order
Post-training · Agents' Last Exam Stage 03 · Featured
Agents' Last Exam
Measures whether agents can finish economically valuable, real industrial work: long-horizon, with a deliverable and a verifiable outcome. Cited in 10 vendors' 2026 model cards. IcarusTech runs 10 production lines for it, authored by industry practitioners and domain experts.
10lines
Production lines
970tasks
In inventory
1,300/ week
Production capacity
22tasks
ALE-100 tasks where both frontier stacks stay below 0.5
Below 0.5: where the training signal is0.5 and above
110–0.3
110.3–0.5
70.5–0.7
230.7–0.9
480.9–1.0
Tasks per score band. Mean best score: Claude Opus 4.8 0.725 · GPT-5.6-luna 0.662
ALE-Legal
Legal
A network of 200+ practising lawyers; 5–10 lawyer-hours and 61–86 scored criteria per task. No frontier model has passed a task in full.
ALE-Finance
Quantitative finance
Real market data: futures, ETFs, A-shares, global indices, FX and crypto spot. An isolated verifier holds the future labels and scores out of sample.
ALE-Video-Editing
Video editing
Real client projects from professional editing studios, with the full project state recorded at every save and export. Not a screen recording.
Post-training · Terminal-Bench 3 & 4 Featured
Terminal-Bench 3 & 4
Real engineering work in the terminal: software, business operations, ML, security and GPU kernels. Each task is calibrated with 8 Claude Opus 5 rollouts, reporting full-solve rate alongside mean score.
553tasks
Delivered
8rollouts
Per task
35%
of tasks never fully solved in 8 tries
0.63
Mean score across the 8 rollouts
Never fully solved: 195 tasksSolved at least once
1950
541–2
493–4
605–6
817
1148
Also
- 49 tasks need a GPU
- Scaffolds: mini-swe-agent 303 tasks, claude-code 250
- In stock: 15K terminal and CLI tasks (Terminal-Bench 2.0 style)
Post-training · Terminal-Bench Science Featured
Terminal-Bench Science
Research tasks on real scientific data: recover the physics from simulation output, reproduce a GPU solver, run science and engineering workflows. Written by 200+ PhDs and postdocs, graded programmatically with no LLM judge.
Scientific understanding · 343
Physics reproduction · 275
Scientific workflow · 600
Collections are scored differently, so scores do not compare across them. GPT-6 Astra serves as the solvability anchor.
3 further tasks are unlabelled.
TBS Hard 136 · new October batch
Understanding physics, 76 tasks: restore up to 13,000 shuffled simulation frames and name every quantity · Physics reproduction, 60 tasks: write a GPU solver from 8 frames, graded on A100 / H100.
19%
of Opus 5.5 attempts reach full score
62%
of tasks Opus 5.5 never fully solves in 4 tries
37%
of tasks no frontier agent fully solves in 6 runs
16hrs
Median expert time to solve
Post-training · new environments Featured
Real codebases and robots
Two new kinds of RL environment: implement a feature inside a real open-source game or app, or write a robot controller in MuJoCo for a real teleoperation task. Both ship in Harbor format, ready to run.
Game & mobile dev RL
Game and mobile development
Starting from the commit before an upstream feature, the agent implements that feature in a real open-source UE5 project or app. Hidden tests score it from 0 to 1.
104tasks
RL tasks
54repos
Upstream open-source repos
5,682tests
Hidden tests
Platform mixTasks · 7 languages
Game dev: 43Mobile dev: 61
Median expert time 5 h (2.5–12 h) · 53% of gpt-6-luna runs reach full score, mean 0.91: mostly solvable, with a hard tail.
Robotic manipulation · MuJoCo
Robot manipulation
From a real teleoperation platform: the agent writes a closed-loop controller and is scored on outcomes only, on randomised layouts it has never seen.
5,000+tasks
Off the shelf
16layouts
Hidden per task
1,000demos
Human demos, all full score
84–100%
Human success rate on these tasks
0.165
gpt-6-luna mean score (0–1)
20 of 160 hidden layouts solved
Robots in the 10-task sampleTasks
Post-training · in stock Ready to ship
More post-training data in stock
Existing stock aligned to well-known benchmarks: RL environments with executable graders, plus expert annotation and rubrics. Ready to ship, with capacity to scale.
Code & SWE
Repository-level SWE
SWE-bench Verified · Pro · Multilingual
30K instancesTerminal & CLI
Terminal-Bench 2.0 · SkillsBench
15K tasksCompetitive programming
LiveCodeBench · FrontierCS · 5K / month
30K problemsLong-horizon multi-round coding
RoadmapBench · EvoCode-Bench
1,200 casesAgents & tools
Tool calling & MCP
τ²-bench · BFCL · MCP Atlas
9,300 tasksDeep search
BrowseComp
50K setsOccupational deliverables
GDPval
4,200 casesEnterprise workspaces
30 workspaces · real app workflows
61K setsReasoning & experts
Expert-authored research & olympiad
Humanity's Last Exam · GPQA
70K problemsMath: exact answers & proof grading
AIME · HMMT · IMO-AnswerBench
10K+ problemsClinical medicine
HealthBench · physician-authored
110K itemsRubrics
PhD math & physics · frontier science · writing
37K rubricsMultimodal
Image reasoning & captions
MMMU-Pro · MathVista
1.25M pairsUI grounding & computer use
OSWorld-Verified · ScreenSpot-Pro
1.8M recordsVideo understanding & captions
Video-MME
200K pairsMultilingual speech annotation
ASR · TTS · full-duplex
30K hoursThe line under each item names the public benchmark it aligns with; figures are approximate. Also: 8,500 designer-spec frontend pages · 20 benchmark reproduction packs
World models Stage 04
Embodied and world-model data
Shipped with control signals, camera calibration and physical ground truth, kept separate by capture platform and viewpoint rather than pooled.
Data typeWhat we haveQuality
Embodied capture & interaction
Real robots355Kepisodes
Cross-embodiment manipulation 300K + egocentric 55K, kept apart by platform and viewpoint.
- Same skill on arms of differing reach and payload
- 30K multimodal episodes: joint states, action chunks, calibration
- Teleop demos to order, failures kept
World-model simulation
Multi-view sim10Khours
Scenes, cameras and dynamic entities generated parametrically, not reprocessed from real video.
- Frame-synced multi-view with camera intrinsics and extrinsics
- Pose, depth and segmentation per frame
- Randomisation ranges declared per scene
Game engine source & assets
UE5 · Unity65titles
Full source for interactive worlds: 5 UE5 games on Steam, 17 Chinese mobile and web games, 43 Unity repos.
- Complete source code and art assets
- Being converted into interactive RL environments
- Can ship alongside the game-dev RL tasks
Original filmed footage
Original capture1,500hours
Filmed to a brief rather than scraped, with the annotation that makes it trainable.
- 1,000 hours of egocentric task video, indexed by activity
- 500 hours of aerial survey in 2K / 4K
- Scripted action capture, 150 hours per month
See also Robot manipulation RL environments (5,000+ in stock, MuJoCo) in Real codebases and robots.
How we build Asset → Task → Benchmark pack → Delivery
How we build data
IcarusTech runs data production as an online pipeline with four layers: assets, tasks, benchmark packs and demand, each holding stock. New assets are pre-produced upward so stock is ready; new requests check stock first and build only the gap.
4layers
Stock at every layer
2,100+
Public benchmarks profiled from 306 model cards
26 × 99
Asset taxonomy: domains × types
0 | 1
Executable grader on every RL task
L1 · AssetsPartner · Asset
Asset intake
Raw data becomes assets with a source, a licence and a health check
- Source: partner portal, in-house collection or public sources
- Record: source, licence scope, volume and price, per asset
- Screening: credentials, personal data, malware and overlap with public corpora; failures held back
- Taxonomy: 26 domains × 99 types
Output Asset cards; corpora and trajectories ship as data products
L2 · TasksFactory
Task production
Production recipes turn each kind of asset into verifiable tasks
- Structure: environment, instructions, grader and reference solution, in Harbor format
- Acceptance: reference solution 1.0, empty and hard-coded submissions 0; offline and reproducible
- Difficulty: measured on current frontier models
- Lineage: every adaptation is a new, traceable version
Output Verifiable RL tasks, plus SFT and eval data
L3 · Benchmark packsFactory
Benchmark fit
Build a task once, then check which benchmarks it fits
- Profiles: 2,100+ public benchmarks: format, grading, domain, difficulty bar
- Fit: each task × benchmark: fits, fits after adaptation, or no fit
- Adaptation: a new format, a new grader or a harder variant
- Leakage: overlap with your named test sets checked before delivery
Output A data pack built for one benchmark
L4 · DemandConsole
Delivery to order
The system drafts a fulfilment plan; a person approves it before anything runs
- Request: benchmark, volume, domain, difficulty, timeline, exclusivity
- Plan: stock + fitting tasks + adaptations + new builds + sourcing, with time estimates
- Pilot: sample tasks first, then volume
- Progress: written back to the order, visible at any time
Output Data pack + acceptance report
What this means for you
1
Fast
Most requests are matched from stock, not built from scratch
2
Verifiable
An offline grader and reference solution ship with every task
3
Traceable
Source and licence record per asset; version lineage per task
4
Stock or exclusive
Off-the-shelf packs, or exclusive data locked after delivery and never resold
Let's build together
Better training data,
built together
Need an off-the-shelf pack, or data and RL environments built to a specific benchmark? Tell us the capability you want to train or evaluate. We build the environment and write the grader.
Off-the-shelf packsBuilt to a benchmarkCustom RL environments