Early accessLLMGR-LLM-03

Agentic Coding Environments

Containerized software and tool-use tasks with automated verifiers.

Request sample Download spec (PDF)Get pricing

Task format: Harbor
Verification: Automated
Reference solutions: Included
Repo languages: 5+

Illustrative rendering of annotation layers.

01Overview

Runnable tasks for training and evaluating coding and tool-use agents: feature work in real open-source repositories, long-horizon engineering projects, terminal tasks, skill-based tasks and multi-service enterprise workflows. Every task ships with its environment, an automated verifier and a reference solution, ready for reinforcement learning.

Modalities

Code
Docker environments
Tests
Tool calls

Use cases

Reinforcement learning for coding and tool-use agents
Agent evaluation beyond public benchmarks
Regression testing of agent harnesses

02Specification

Task families

Repository tasks: Feature requests in real open-source repositories (Python, TypeScript, JavaScript, Go, Rust), graded by fail-to-pass and pass-to-pass tests
Long-horizon engineering: Reimplementation and performance work with byte-exact or benchmark-based verifiers
Terminal: Linux command-line and scientific-computing tasks in containers
Skills: Tasks solved with packaged agent skills and checked by output tests
Enterprise services: Workflows across self-hosted GitLab, ownCloud, Plane and Rocket.Chat for fictional companies, scored against checkpoints
MCP tool use: Multi-step tasks over MCP servers with rule-based ground truth

Each task includes

Instruction: Task statement in English (some skill tasks in Chinese)
Environment: Dockerfile and repository snapshot or seeded services
Verifier: Tests, grader script or checkpoint rubric
Reference solution: Patch or script that passes the verifier

Provenance

Tasks: Newly written; not copies of public benchmark tasks
Upstream code: Open-source repositories under their original licenses
Model trajectories: Not included
Volume: Confirmed per order

03Schema

Field	Type	Description
task.toml	toml	Task metadata, environment and timeouts.
instruction.md	markdown	Task given to the agent.
environment/	directory	Dockerfile and task inputs.
tests/	directory	Verifier that scores the agent's result.
solution/	directory	Reference solution.

04File structure

Directory layout

agentic-coding-environments/  # indicative; final layout confirmed per delivery
└── tasks/<task_id>/
    ├── task.toml
    ├── instruction.md
    ├── environment/
    │   └── Dockerfile
    ├── tests/
    │   └── test.sh
    └── solution/
        └── solve.sh

05Sample

Sample on request.

Runnable tasks from each family with verifiers and reference solutions

Size: On request
Formats: Harbor tasks, Docker

Request sample Download spec (PDF)

06License

Graviti Commercial Data License

Commercial training and evaluation rights. Upstream open-source code keeps its own license.

How we license data →

07FAQ

Questions about Agentic Coding Environments

What is the Agentic Coding Environments dataset?

Agentic Coding Environments (GR-LLM-03) is a LLM training dataset from Graviti. Runnable tasks for training and evaluating coding and tool-use agents: feature work in real open-source repositories, long-horizon engineering projects, terminal tasks, skill-based tasks and multi-service enterprise workflows. Every task ships with its environment, an automated verifier and a reference solution, ready for reinforcement learning.

What is Agentic Coding Environments used for?

Typical uses: reinforcement learning for coding and tool-use agents, agent evaluation beyond public benchmarks and regression testing of agent harnesses.

What modalities and annotations does Agentic Coding Environments include?

Agentic Coding Environments includes Code, Docker environments, Tests and Tool calls. Its schema has the fields task.toml, instruction.md, environment/, tests/ and solution/.