Request access to the Revelo Terminal-Bench 4.0 task samples
Revelo reviews each request, usually within one business day. These samples are for evaluation by your team only. Please do not use them to train models, and please do not share them outside your organization. Please use your work email. We may decline requests we cannot verify.
Log in or Sign Up to review the conditions and access this dataset content.
Revelo Terminal-Bench 4.0 tasks: samples
Dataset Description
20 sample tasks from Revelo's set of 5,000+ terminal tasks for AI agents.
Each task puts an agent in a Linux container with existing code or data and a written request. The agent must build, repair or finish working software, for example a deadlock model in VHDL, a margin analyzer in Rust or a geological model in C++. A hidden test suite then checks the result in a separate container. The agent never sees the tests, and partial work does not pass.
The tasks cover hardware, science, security, finance and software engineering. Each task gives the agent 4 to 8 hours.
For evaluation only. Please do not use these samples to train models.
The two sample sets
| Set | Tasks | Claude Sonnet 5.5 | GPT-6 Luna | Hy4 Preview |
|---|---|---|---|---|
hard/: frontier models fail most runs |
10 | 13 of 80 runs pass (16%) | 13 of 100 runs pass (13%) | 3 of 50 runs pass (6%) |
solved/: at least one frontier model solves every run |
10 | 80 of 80 runs pass (100%) | 55 of 100 runs pass (55%) | 22 of 50 runs pass (44%) |
A run passes when the held-out verifier gives a reward of 1.
Dataset Structure
Each task is a Harbor task with this layout:
<task>/
├── harbor_task/
│ ├── instruction.md # the task the agent receives
│ ├── task.toml # metadata, resource limits, verifier settings
│ ├── environment/ # Dockerfile and starting workspace
│ ├── tests/ # held-out verifier, runs in a separate container
│ └── solution/ # expert reference solution
└── rollouts/
└── <model>/<run>/ # full trajectory, submission, verifier output, reward
- A senior engineer authors every task. Two independent reviewers check it.
- Each task includes the expert's reference solution.
task.tomlexplains why the task is hard, how the reference solution works, and what the verifier checks.
Tasks
Hard set
| Task | Domain | Claude Sonnet 5.5 | GPT-6 Luna | Hy4 Preview |
|---|---|---|---|---|
| chi-scc-witness | Hardware / RTL | 3/8 | 0/10 | 0/5 |
| charger-rk4-ribbons | Science / Physics | 1/8 | 3/10 | 0/5 |
| citygml-barycentric-meshing | Software / Algorithms | 2/8 | 1/10 | 0/5 |
| coseismic-trapezoid-clipping | Science / Earth | 1/8 | 2/10 | 1/5 |
| flatzinc-blossom-transform | Software / Algorithms | 0/8 | 2/10 | 1/5 |
| imputation-release-audit | Science / Biology | 0/8 | 0/10 | 1/5 |
| margin-transition-reconciliation | Software / Data engineering | 1/8 | 0/10 | 0/5 |
| roster-causal-frontier | Software / Frontend | 2/8 | 3/10 | 0/5 |
| session-joint-transport | Science / Math | 0/8 | 1/10 | 0/5 |
| trial-witness-cover | Software / Algorithms | 3/8 | 1/10 | 0/5 |
Solved set
| Task | Domain | Claude Sonnet 5.5 | GPT-6 Luna | Hy4 Preview |
|---|---|---|---|---|
| atlas-liquidity-sweep | Operations / Finance | 8/8 | 9/10 | 4/5 |
| bridge-causal-certificate | Security / Forensics | 8/8 | 1/10 | 2/5 |
| canonical-phrase-dag | Software / Algorithms | 8/8 | 8/10 | 1/5 |
| hl7-panel-ownership | Software / Systems | 8/8 | 2/10 | 0/5 |
| lectionary-layout-dp | Software / Algorithms | 8/8 | 8/10 | 3/5 |
| lexical-history-incidence | Software / Database engineering | 8/8 | 6/10 | 1/5 |
| meaning-history-minloss | Software / Algorithms | 8/8 | 5/10 | 3/5 |
| mlkem-anchored-moments | Security / Cryptography | 8/8 | 8/10 | 4/5 |
| report-lock-compositor | Software / Algorithms | 8/8 | 1/10 | 1/5 |
| rna-velocity-cdf | Science / Biology | 8/8 | 7/10 | 3/5 |
Usage
From a task's harbor_task/ folder, with Harbor installed:
harbor run --path . --agent oracle # reference solution: reward 1
harbor run --path . --agent nop # untouched workspace: reward 0
License
Revelo builds these tasks in the Terminal-Bench 4.0 format. License all 5,000+ as is, or have us extend them with new domains or harder tasks. Request the full set.
These samples are for evaluation by your organization only. Please do not use them to train models, and please do not share them outside your organization. The canary string in each file marks this content as benchmark data that must stay out of training corpora.
Contact
- Downloads last month
- 294