Community¶
AOBench is a small project that would like to be a useful one. That depends on people outside Bologna running it, disagreeing with it, and fixing it.
Where things happen¶
| Space | For |
|---|---|
| Discussions | Questions, ideas, results you got, "is this a bug or am I holding it wrong" |
| Issues | Bugs, concrete proposals, task and environment contributions |
| Good first issues | Small, well-specified work with an honest time estimate |
| Start here (#20) | A map of every open area to a first task |
| Security advisories | Vulnerabilities — please not the public tracker |
Questions are welcome and expected. If the docs did not answer something, that is a documentation bug, and asking is how we find out.
What we most need¶
Ordered by how much difference it makes, not by how hard it is:
- An environment snapshot from a real facility. Twenty-three of twenty-nine environments are synthetic. One sanitised window from a cluster that is not Marconi100 would improve this benchmark more than months of our own work. See adding an environment.
- Tasks. Corpus breadth is what limits what AOBench can measure — especially
DATAandPERF, which are thin because neither maintainer is a storage specialist. See adding a task. - Results. Run it on your agent and publish the numbers, including bad ones. See the leaderboard.
- Adapters. Every new adapter makes a whole class of agent evaluable by everyone.
- Reports that we got something wrong. Especially a wrong gold answer — those are the most damaging bug a benchmark can have, and reporting one is a favour, not a criticism.
- Code, docs, tests. The usual, always welcome.
What you can expect from us¶
- A first response within 3 working days, even if it is only "seen, I'll look properly on Friday".
- If a PR goes quiet for over a week, ping it. That is our failure, not rudeness on your part.
- Credit. Every merged contribution earns a line in AUTHORS.md and a mention in the release notes. Substantial corpus or methodological work may warrant paper co-authorship — if you think that applies to you, say so.
- Public reasoning. Decisions are made in issues, with the argument written down. See GOVERNANCE.md.
What helps us¶
- PRs under ~300 changed lines. Bug fixes, docs, tests, examples, and new CLI flags need no prior discussion — just send them.
- Open an issue first only if you are changing the task schema, the scoring weights, the RBAC model, or a public CLI signature.
- You do not need HPC access. The whole benchmark runs against frozen snapshots on a laptop, and
direct_qaneeds no API key.
For researchers¶
RESEARCH.md lists sixteen open questions AOBench can now answer empirically. Several are good MSc or PhD projects; a couple are a good afternoon. We are happy for you to publish first-author, and co-supervision is possible for students at any institution.
Negative results are welcome here. "We tried X and it did not help" is a real contribution to a young field, and this project will link it rather than bury it.
Code of conduct¶
The Code of Conduct applies everywhere. Short version: disagree about the benchmark as much as you like, and be decent to the person you are disagreeing with.