| --- |
| license: apache-2.0 |
| tags: [openenv, rl-environment, agents, terminal, tracelite] |
| pretty_name: TraceLite EnvBench (OpenEnv) |
| --- |
| |
| # TraceLite EnvBench (OpenEnv) |
|
|
| Five realistic ops and data jobs from TraceLite AI's EnvBench. Each task reads like an |
| ordinary ticket ("add a hall column to the sensor statistics", "put the card scheme on every |
| settlement line"). What decides success is a property of the environment that the prompt never |
| mentions: line endings, a missing toolchain, inherited shell or compiler settings, a non-English locale. |
|
|
| **What it teaches.** Check the conditions you run in before trusting a solution that looks right: |
| read the docs and the actual output, test on the real data, verify the end state. |
|
|
| **How the reward works.** The agent has one tool (`{"command": "<bash>"}`) and ends with |
| `{"submit": true}`. A hidden deterministic grader then runs the delivered job the way production would and |
| checks files and service state (output files, exit codes, alerts raised). Reward is 1.0 or 0.0. For every task the reference solution scores 1, |
| submitting without doing anything scores 0, and the "idiomatic" solution most engineers would write |
| scores 0. |
|
|
| Task IDs: see `tasks.jsonl`. Difficulty was calibrated on the arena's base model (Qwen3.8-27B, 6 attempts per task): each task is solved in 1 to 4 of 6 attempts. |
|
|
| Built by [TraceLite AI](https://tracelite.ai). Graders and setup data ship inside the image in obfuscated |
| form and are staged only after the agent submits. |
|
|
